2b9a1083b5e3c8690db7c6199c92a78f1d8a6e83
1015 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Brummel
|
e25580e3a3 |
test(cutover): live accept/reject corpus for the #55 ownership surface
Wires the #55-cutover fieldtest into a durable regression net. The
fieldtest exercised the post-cutover ParamMode={Own,Borrow} surface as a
downstream LLM author would and produced 12 fixtures + a report; left as
loose files they would be dead fixtures that drift. This commits them as
a protected property.
crates/ail/tests/cut55_cutover_surface.rs — one table-driven test running
each fixture through `ail check` and asserting accept (exit 0) or reject
at the right pipeline stage. Reject rows key on the bracketed diagnostic
CODE (consume-while-borrowed, borrow-over-value, borrow-return-not-permitted,
surface-parse-error), NOT message text, so the test is decoupled from the
diagnostic-wording improvement tracked separately.
examples/fieldtest/cut55_*.ail — the corpus. cut55_2c is the #58 repro
(now correctly rejected). cut55_2b was misanalysed by the fieldtest as a
consume-while-borrowed reject; in fact bump's recursive param is
(borrow List) so the tail is only borrowed, never consumed into an own
slot — `ail check` correctly accepts it. Renamed to
cut55_2b_borrow_traversal_clean and recast as the #58 false-positive
guard (a borrow-position use of a heap sub-binder must stay accepted).
cut55_1b similarly does not fire over-strict-mode (its recursive call
consumes the tail, so the lint's consume_count==0 premise fails) — comment
corrected; it is a plain accept. cut55_1c is the genuine over-strict-mode
example (accepts exit 0, emits the warning).
docs/specs/0065-eliminate-implicit-mode fieldtest report — moved 2c to
the reject set, added the 2b false-positive-guard section, marked the
double-free spec_gap RESOLVED (it shipped as the #58 fix).
Relates to #55.
|
||
|
Brummel
|
19757480b7 |
audit(0064/cutover): record shipped drop machinery + retire dead desugar arm
Cycle-close tidy for the #55 cutover (its audit drift-resolution). Architect drift review found the design ledger out of step with the codegen that the cutover shipped, plus two debt items; regression green (733 passed / 0 failed across the workspace). design/contracts/0008-memory-model.md — the Codegen contract described the drop-emission *gates* but not the drop machinery the cutover landed. Added three subsections recording the present state (honesty rule): - Per-monomorph drop fns for polymorphic ADTs (crates/ailang-codegen/ src/dropmono.rs: collect_drop_monos / DropAdtMeta, the suffix scheme, value-field-skip on the substituted field type). Ratified by alloc_rc_value_type_field_is_not_rc_dec_dropped. - String-literal rep promotion at owned drop sites (StrRep::Static→Heap on a Str literal flowing into an Own slot the callee drops; gated strictly on Own). Ratified by own_str_literal_arg_is_dropped_exactly_once. - Same-constructor arm grouping in match desugar (build_chain / build_ctor_group bind ctor fields once). Ratified by lit_pat_ctor_tail_drop_single_drop + lit_pat_nil_scrutinee_single_drop. crates/ailang-core/src/desugar.rs — build_chain routes every Ctor-head arm (including a length-1 run) through build_ctor_group, so desugar_one_arm's Pattern::Ctor arm was unreachable dead code duplicating the bind-once lowering. Replaced it with unreachable! naming the invariant, and corrected two stale doc comments (the module-header step 4 and the desugar_one_arm doc still described it as handling ctor arms, and mislabeled the Lit lowering as Term::Match — it is Term::If). Full workspace stayed green, confirming the arm was dead. crates/ailang-check/src/linearity.rs — renamed test implicit_fn_is_exempt → own_param_consumed_once_is_clean. Post-0062 there is no Implicit and no exemption; the test pins the ordinary single-Own-consume clean path. Did the desugar + linearity edits inline rather than via an agent: I had already loaded build_chain/build_ctor_group/desugar_one_arm to prove the arm dead, and a sub-agent would have redone the same reading. |
||
|
Brummel
|
b129af1363 |
fix(check): inherit borrow on heap sub-binders of a borrowed scrutinee (#58)
The strict linearity check accepted consuming a heap-typed pattern sub-binder extracted from a `(borrow)`-mode match scrutinee. The sub-binder aliases the caller-owned interior, so moving it into an `(own)` slot double-frees at runtime (SIGSEGV) — `ail check` returned exit 0, `ail build`+run crashed. A pre-existing hole the corpus never exercised; the #55 cutover's universal activation of the strict check exposed it (sibling to the #57 / spec 0064 class-2 let-alias-of-borrow hardening — this is the pattern-sub-binder analog). Root: `walk_arm` seeded every pattern sub-binder with `borrow_count = 0` (freely consumable), bypassing the `Borrow`-param seeding in `check_fn` that starts a borrowed param at `borrow_count = 1`. The whole-param consume was caught (cut55_2d, correctly rejected); the moved-out heap sub-binder was not. Fix: in `Term::Match`, resolve the scrutinee through `resolve_alias` and record whether it is a borrowed binder; thread that into `walk_arm`, which now seeds a heap-typed (`!is_value`) sub-binder of a borrowed scrutinee with `borrow_count = 1`. Value-typed sub-binders (e.g. the `Int` head) stay at `0` — copied out, no heap alias. Nested matches fall out for free: the inner match re-derives borrowed status from the borrow-carrying sub-binder. Rejected alternative: forcing fieldtest fixture cut55_2b to also reject. 2b is a clean borrow-traversal — `bump`'s recursive param is `(borrow)`, so the tail `t` is only borrowed, never consumed into an `(own)` slot. Distorting the fix to reject it would be unsound (false positive on legitimate borrow-rebuild). The fixture's own comment misdescribed its recursive param as `(own)`; the code says `(borrow)`. Verification: - RED test linearity::tests::consume_heap_sub_binder_of_borrow_scrutinee_is_rejected: was left:0 right:1 (check returned []), now green. - Full workspace: 732 passed / 0 failed (was 731; +1 the new test). No committed corpus newly rejected by the sharper check. - fieldtest repro cut55_2c: exit 0 (wrongly accepted) -> exit 1 (consume-while-borrowed on `t`). closes #58 |
||
|
Brummel
|
76b21c00eb |
feat(lang): eliminate the Implicit ownership default — totality + the drop-soundness it demasks (#55)
Deletes `ParamMode::Implicit`. `ParamMode` is now `{Own, Borrow}`:
every fn-type slot on every signature carries an explicit `own` or
`borrow`, no defaulted position survives anywhere (model 0008 §2,
spec 0062). The parser rejects a bare fn-type slot; `borrow-return`
and `borrow-over-value` reject at the signature; the corpus is
migrated to minimal-ownership modes (consumed ⇒ own, read-only-heap
⇒ borrow, value ⇒ trivial-own). The documented `Implicit`-ret-mode
leak is fixed: an owned heap return now drops exactly once (live=0,
acceptance criterion 5).
This was the easy half. Removing the default ACTIVATED a family of
drop paths that `Implicit` had silently skipped — the pre-cutover
language was leaking (and in places mis-dropping) here rather than
crashing, because an Implicit scrutinee turned the drop off. Making
the modes explicit (Own) turned those paths on and exposed two
latent-bug clusters, all fixed RED-first as part of this cutover:
Drop-soundness family (four legs):
A. lit-sub-pattern double-free — the desugar re-matched the same
owned scrutinee in the lit fall-through; fixed by grouping
consecutive same-ctor arms into one match (bind fields once),
in ailang-core desugar.
B. Cons-husk leak on non-tail arm bodies — the lit-sub-pattern
desugar rebound the owned scrutinee via `Let $mp = xs`, which
bumped consume_count and suppressed the existing fn-return
partial_drop. Fixed by not rebinding a bare-Var scrutinee
(one husk-freeing mechanism, not two).
C. polymorphic `drop_<T>` rc_dec'd monomorphised value fields —
the per-ADT drop fn was emitted once from the polymorphic
TypeDef, defaulting type-var fields to ptr and rc_dec'ing
inline Ints (segfault). Fixed with per-monomorph drop
functions (new ailang-codegen::dropmono): the drop set is
collected from the lowered MIR, value-type fields are skipped,
heap fields still freed once; monomorphic-concrete ADTs keep
their byte-identical un-suffixed drop symbol.
D. static Str literal passed to an `(own Str)` param — the
literal lowers to a header-less rodata constant; the callee's
now-active rc_dec read its length field as a refcount and
freed a static address (segfault). Fixed with the missing
fourth StrRep::Static→Heap promotion in lower_to_mir's App arm,
gated on Own mode (borrow args stay static, no regression).
over-strict-mode lint over-fired: it suggested `(borrow V)` for
value-typed params (which `borrow-over-value` rejects — own is the
only legal mode there) and fired on `(intrinsic)` bodies (whose
consumption the linearity walk cannot observe). Tightened to skip
both; contract 0008 updated to the narrowed firing scope.
Irreversible step — canonical-form hash reset (model 0008 §6,
acceptance criterion 6). Every signature now carries explicit modes,
so the hashable canonical JSON changed for every module. RATIFY:
the corpus-wide hash-pin reset (hash_pin, prelude_module_hash_pin,
mono_hash_stability, eq_ord_e2e, embed_export_hash_stable, the
ct4/iter*/loop_recur schema-extension pins) and the list ir_snapshot
golden were regenerated once, deliberately, as the intended one-time
consequence of removing the mode elision from the canonical form —
not a regression. Each regenerated hash verified deterministic across
two runs.
Also fixes a pre-existing latent failure surfaced by the verification
gate, unrelated to this cutover: the `every_contract_names_a_resolvable_
ratifying_test` resolver (design_index_pin) could not resolve the
" + " dual-link ratifying-test form (`uniqueness.rs + linearity.rs`)
that the #57 audit-close (
|
||
|
Brummel
|
05c3c018de |
fix(check): branch Float NoInstance addendum on method name (#25)
The Float-aware `NoInstance` addendum emitted the same hint for every Eq/Ord method at Float: "use float_eq / float_lt (and siblings ...)". For `eq`/`ne`/`lt`/`le`/`gt`/`ge` that is correct — each has a drop-in boolean `float_*` sibling. For `compare` it is incomplete and misleading: `compare` returns three-way `Ordering`, but no `float_compare` ships (a deliberate non-ship, docs/specs/2026-05-20-operator-routing-eq-ord.md). An LLM author who reached for `compare` to get three-way ordering was pointed at boolean siblings as if a drop-in existed, and left guessing whether to reconstruct it themselves. Fix: branch the addendum on the called method. The boolean arm keeps the existing wording; the `compare` arm instead names the deliberate `float_compare` omission and points at building a three-way `Ordering` from `float_lt` + `float_eq`. One added branch on the existing `method` field (already carried by `CheckError::NoInstance`), no new machinery; the structured `ctx` is untouched. Out of scope (the issue's deliberate non-ship): shipping a `float_compare` fn. The friction was that the diagnostic did not acknowledge the omission — now it does. RED-first, inline: a one-branch diagnostic-text change at a single site whose surrounding code was already loaded while working the adjacent #52 fix this session (CLAUDE.md "context already loaded" carve-out). TDD discipline preserved by the new RED test `compare_at_float_names_deliberate_no_float_compare` over a new minimal fixture `examples/compare_float_noinstance.ail` (`compare` at Float); it failed RED on the old shared wording and passes GREEN on the branch. The pre-existing `eq_at_float_fires_float_aware_noinstance` still passes (eq arm unchanged). Verification: full ailang-check lib suite (121) + full ail test suite green; live CLI renders both arms correctly. closes #25 |
||
|
Brummel
|
d8be95dd44 |
fix(check): render restricted-set diagnostic in AILang set syntax (#52)
The `param-not-in-restricted-set` diagnostic rendered its allowed
element set with `{allowed:?}` — a Rust `{:?}` debug print of the
`Vec<String>`, producing `["Bool", "Float", "Int"]` with quotes,
brackets, and Rust list shape. AILang's audience is an LLM author, so a
language-level diagnostic leaking Rust formatting into its surface is a
quality defect.
Fix: render the set as `{Bool, Float, Int}` via a thiserror trailing-arg
`.allowed.join(", ")` wrapped in literal braces. The set is alphabetical
and stable (unchanged — the source is a sorted BTreeSet projection).
Design decision (set notation, not a union): the allowed set is a
membership set ("one of"), and AILang has no union-type surface syntax,
so brace set-notation `{A, B, C}` reads naturally without implying a
type-level union the language does not have. The structured JSON detail
(CheckError::detail, the `allowed` array) is left untouched — it is
machine-readable and correctly an array, not the cosmetic surface.
RED-first, inline: the bug is a one-line Display-string change at a
single site (lib.rs ~760). The exact lines were already loaded while
assessing the /boss queue, so a debugger subagent would only redo the
same reading (CLAUDE.md "When NOT to delegate: context already loaded").
TDD discipline is preserved by the new RED test
`param_in_reject_message_renders_allowed_set_in_ailang_syntax`, which
asserts the AILang set form and the absence of Rust quotes/brackets;
it failed RED on the old `{allowed:?}` and passes GREEN on the fix. The
pre-existing pin `param_in_reject_message_names_allowed_set` only
asserted substring presence, so it did not pin (or block) the format.
Verification: full ailang-check suite green (121 lib tests + integration
suites); live CLI repro from the issue now renders
`(allowed: one of {Bool, Float, Int})`.
closes #52
|
||
|
Brummel
|
9848d4b2df |
docs(design): explore explicit RC as opt-in, not the floor (model 0009)
Records the runtime-axis dual of ownership totality (model 0008). 0008 totalises the *annotation* (own/borrow everywhere, no Implicit) but keeps universal RC as the floor; 0009 asks the question totality makes askable — must the runtime discipline stay universal, or can it follow the type? Principle: RC is authored, never universal. Default `box` (affine, refcount-free, deterministic scope-drop ≈ Box); opt-in `rc` (refcounted, explicit `share` ≈ Rc). Motivated by the maintenance record: ~1/3 of recent commits touch RC/ownership/drop, no bug was ever in the RC runtime itself — all lived in the inference (uniqueness-elision, is_rc_heap_allocated, mode-through-subst) that exists only to reconcile affine ownership with universal RC and strip the RC it over-charges. Soundness leans on existing law: the 0015 DAG constraints + linear-by- default already draw the "needs-a-count" partition; `rc` only names it. No cycle collector reappears (rc graphs stay acyclic). No `&mut` mode: in-place mutation rides consume-and-return (move, not alias), so constraint 3 (immutable-once-constructed) holds — `mut` is out by principle. §5 enumerates what leaves the compiler (uniqueness-elision, is_rc_heap_allocated term-shape predicate, over-strict lint, reuse-as, escape-clawback) and what stays (the affine/linearity check, already present). Status-stamped NOT current state, per the honesty rule and the 0008 precedent; INDEX.md left untouched (exploration whitepapers are not on the canonical spine until adopted). Composes with 0008 regime A, which already makes transitive box-drop sound. §8 isolates the open questions (closure capture move/share, clone-vs-share split, coupling to 0008's escape pass, keyword spelling, sharing-heavy workloads). |
||
|
Brummel
|
dfdc65f21f |
audit: spec 0064 cycle close — drift-clean after two ledger fixes (#57)
Cycle-close audit for spec 0064 (the #57 hardening of the linearity analysis, the #55-cutover precondition; four iterations |
||
|
Brummel
|
2769e50a35 |
fix(examples): destructure partition_eithers result once (#57, 0064 iter 4)
Final iteration of spec 0064 (the #55-cutover hardening, #57). Closes class 4, the one genuine corpus bug among the four (not a false positive): std_either_list.partition_eithers projected both (app Pair.fst rest) and (app Pair.snd rest) from one owned `rest`. Pair.fst/snd are own-param projections that move a field out, so each consumes `rest` -> a genuine double-consume that universal linearity activation (#55) would correctly reject. The fix is a body rewrite (destructure `rest` once via (match rest (case (pat-ctor MkPair ls rs) …)) then dispatch on the head), NOT a check change -- the check is right. Latent until activation: partition_eithers' param is a bare (Implicit) slot today, so the check is skipped on it -- no RED->GREEN flip on the real fixture. The rewrite is a pre-emptive corpus fix. Its correctness rests on the unchanged 2 3 2 3 E2E output (std_either_list_demo, a semantically-identical partition) and on the new c4_double_consume must-stay-RED fixture, which -- with explicit modes, so the check is active today -- demonstrates the double-projection shape IS rejected. - examples/std_either_list.ail: partition_eithers Cons arm rewritten; doc updated (destructure-once, no longer "projections"). - examples/c4_double_consume.ail: must-stay-RED pin (the rejected shape), folded into a generalised harden_ownership_heap_double_consume_still_errors loop over {real_consume, c4_double_consume}. - examples/c4_rewrite.ail: stays-clean pin (the accepted destructure-once shape), added to harden_ownership_false_positives_are_clean. No hash-pin on std_either_list (verified); the e2e 2 3 2 3 output pin is the content pin and is preserved. No check/schema/codegen change. Verified: std_either_list checks clean; demo 2 3 2 3; both harden assertions green; cargo test --workspace green; bench/check.py 34/34, compile_check.py 24/24 stable. This is the last of the four #57 hardening classes. Spec 0064 is now code-complete: classes 1 (local fn-param modes), 2 (let-alias redirect), 3 (value-typed let-binder), and 4 (this) all shipped. The cycle is ready for audit; the #55 cutover precondition is met. refs #57 |
||
|
Brummel
|
dea93a99fc |
plan: 0064 iter 4 (final) — partition_eithers double-consume rewrite (class 4) (#57)
Final iteration of spec 0064. Closes class 4, the one genuine corpus bug (not a false positive): partition_eithers projects both (app Pair.fst rest) and (app Pair.snd rest) from one owned rest; fst/snd are own-param projections that move a field out, so rest is consumed twice. Universal linearity activation (#55) would reject it. Fix is a body rewrite (destructure rest once via match), not a check change. Latent-bug nature: partition_eithers has a bare (Implicit) param slot today, so the check is skipped on it -- no RED->GREEN flip on the real fixture. The rewrite's correctness is proven by the unchanged 2 3 2 3 E2E output (std_either_list_demo) and by the explicit-mode c4_double_consume must-stay-RED fixture (the double-projection shape IS rejected, verifiable today). Orchestrator verified the rewrite this session: ail check std_either_list -> exit 0; ail run the demo -> 2 3 2 3 (applied, ran, reverted for the implementer to re-apply). Two tasks: Task 1 rewrites partition_eithers + gates on the unchanged E2E output; Task 2 adds c4_double_consume (must-stay-RED, folded into a generalised harden_ownership_heap_double_consume_still_errors loop) and c4_rewrite (stays-clean, added to the false-positives-clean array). plan-recon confirmed no hash-pin on std_either_list (only the e2e 2 3 2 3 output pin, preserved). No check/schema/codegen change. This is the last of the four #57 hardening classes; after it lands, spec 0064 is code-complete and the cycle is ready for audit. refs #57 |
||
|
Brummel
|
be259f9d46 |
feat(check): let-alias borrow propagation via alias-redirect (#57, 0064 iter 3)
Third iteration of spec 0064 (the #55-cutover hardening, #57). Closes false-positive class 2: Term::Let walked its value in Position::Consume, so (let a t (match a …)) over a borrow-mode t consumed t at the binding and tripped consume-while-borrowed -- a let-alias of a borrowed value read in a borrow position is not a consume. Fix: the Checker gains a scoped aliases: HashMap<String,String>. A (let a t body) whose value is a bare Var resolving to a tracked binder records a -> root(t) for the body scope and skips the consume walk of the value. A new resolve_alias helper maps a name to its root, preferring a real binder at each chain step -- so an inner binder named `a` (nested let, pattern binder, lam param) automatically shadows the alias with NO change to with_binder/walk_arm/Lam (the shadowing is resolved centrally in the resolver, not at three scattered introduction sites). Every binder-state lookup keyed by name resolves through the map first: use_var, the App borrow bump + decrement (pushing the resolved root to `lent` so the decrement matches), callee_arg_modes, and the reuse-as source. Design calls (orchestrator): - Alias REDIRECT, not state clone (spec scope decision 3): consuming the alias marks the single root consumed, so a real double-consume through an alias is still caught. A clone would track two independent binders and miss it (unsound). Pinned by the new let_alias_of_owned_double_consume_still_errors unit test: (let a p0 (seq a p0)) over an own heap p0 still fires use-after-consume. - resolve_alias prefers binders -> shadowing needs no edits to the binder-introduction sites. Lower-risk than clearing aliases at each. - reuse-as source (linearity.rs:663) is resolved through aliases too: it is the 4th binder-state-reading site; the spec's Fix 3 named three illustratively. Without it, (reuse-as <alias> …) would mis-fire reuse-as-source-not-bare-var on what IS a Var. Faithful completion of Fix 3. The diagnostic now reports the resolved root name (the actual consumed resource), not the surface alias. Closes the design/contracts/0008-memory-model.md let-alias carve-out (was "has not shipped yet") and extends its Ratified-by trailer to name linearity.rs, where the propagation now lives -- a contract change riding with the feature that forces it. RED-first: examples/c2_let_alias.ail added to harden_ownership_false_positives_are_clean (RED: consume-while-borrowed on count's t); GREEN after. Verified: c2 RED->GREEN; 120 linearity unit tests green; the soundness test fires; harden false-positive + heap double-consume guards green; cargo test --workspace green; bench/check.py 34/34, compile_check.py 24/24 stable. Diagnostic-only; no schema/type/codegen change. Class 4 (partition_eithers body rewrite) remains for the final iteration of spec 0064. refs #57 |
||
|
Brummel
|
9c4d195806 |
plan: 0064 iter 3 — let-alias borrow propagation (class 2) (#57)
Third iteration of spec 0064. Closes false-positive class 2: Term::Let walks its value in Position::Consume, so (let a t (match a …)) over a borrow-mode t consumes t at the binding -> false consume-while-borrowed. Fix: Checker gains a scoped aliases: HashMap<String,String>. A (let a t body) whose value is a bare Var resolving to a tracked binder records a -> root(t) for the body scope and skips the consume walk. A new resolve_alias helper maps a name to its root, PREFERRING a real binder at each chain step -- so an inner binder named `a` (nested let, pattern binder, lam param) automatically shadows the alias with NO change to with_binder/walk_arm/Lam (the shadowing is handled centrally in the resolver, not at three scattered introduction sites). Every binder-state lookup keyed by name resolves through the map first: use_var, the App borrow bump + decrement, callee_arg_modes, and the reuse-as source. Design calls (orchestrator): - Alias REDIRECT, not state clone (spec scope decision 3): consuming the alias marks the single root consumed, so a real double-consume through an alias is still caught -- a clone would miss it (unsound). Pinned by the let_alias_of_owned_double_consume_still_errors unit test. - resolve_alias prefers binders -> shadowing needs no edits to the three binder-introduction sites (simpler, lower-risk than clearing aliases at each). - reuse-as source (linearity.rs:663) is the 4th binder-state-reading site; the spec's Fix 3 named three illustratively. Resolving there too avoids a spurious reuse-as-source-not-bare-var on an aliased Var. Faithful completion of Fix 3, documented as such. Three tasks, RED-first: Task 1 adds examples/c2_let_alias.ail to the harden list (RED: consume-while-borrowed on count's t); Task 2 is the alias mechanism (GREEN) + the soundness unit test; Task 3 closes the design/contracts/0008-memory-model.md:340 let-alias carve-out (was "has not shipped yet") and extends the Ratified-by trailer to name linearity.rs. Diagnostic-only; no schema/type change. plan-recon mapped current post-iter-2 line numbers; parse gate fired on the inlined fixture (RED confirmed). Class 4 (partition_eithers rewrite) remains for the final iteration. refs #57 |
||
|
Brummel
|
cc5991eeb3 |
feat(check): resolve local function-typed binder modes (#57, 0064 iter 2)
Second iteration of spec 0064 (the #55-cutover hardening, #57). Closes false-positive class 1: applying a local function-typed binder -- a HOF predicate param like std_list.filter's `p` -- treated its arguments as Consume because callee_arg_modes resolved only the global symbol table (locals returned an empty mode vec). So `(app p h)` with `p` a local (borrow ...) predicate consumed the heap element `h`, and reusing `h` in the kept Cons false-fired use-after-consume. Fix: BinderState gains fn_param_modes: Option<Vec<ParamMode>>, seeded at all three function-typed binder introduction sites via a new fn_modes_of helper (strip_forall + Type::Fn { param_modes }, None for non-fn / empty-modes so the caller falls through to globals): params and lam-params from their signature type (param_tys), let-binders from the LetBinderTypes table built in iter 1. callee_arg_modes reads a local Var callee's fn_param_modes before the global lookup (a local binder shadows a global -- correct lexical scope); the existing App arg-walk maps a Borrow mode to a borrow walk unchanged, so the fix is confined to the seeding + the one local-precedence read. Scope call (orchestrator): all three seeding sites, not just the param path that unblocks filter -- the point of this hardening cycle is to leave no latent class (the #57-from-0063-deferral lesson), and the extraction is uniform with the table already built. RED-first: examples/c1_local_hof.ail added to harden_ownership_false_positives_are_clean (RED: use-after-consume on filter_box's h) before the fix; GREEN after. Two in-source unit tests pin both seeding paths (param via signature, let via a hand-built table). Verified: c1 RED->GREEN; 119 linearity unit tests green; harden_ownership_false_positives_are_clean green (now c1 + c3); the heap-double-consume must-stay-RED guard still fires; cargo test --workspace green; bench/check.py 34/34, compile_check.py 24/24 stable. Diagnostic-only; no schema/type/codegen change. The stale `// Locals never carry fn-types in 18c.2` comment in callee_arg_modes (and the fn's doc header) is rewritten to the accurate local-precedence behaviour -- it sat in the replaced lines and directly contradicted the new path. Classes 2 (let-alias redirect) and 4 (partition_eithers rewrite) remain for later iterations of spec 0064. refs #57 |
||
|
Brummel
|
2ad58ada17 |
plan: 0064 iter 2 — local function-typed binder modes (class 1) (#57)
Second iteration of spec 0064. Closes false-positive class 1: applying
a local function-typed binder (a HOF predicate param like
std_list.filter's `p`) treats its args as Consume because
callee_arg_modes resolves only globals -- so a heap arg reused after
(app p h) false-fires use-after-consume.
Fix: BinderState gains fn_param_modes: Option<Vec<ParamMode>>, seeded
at all three function-typed binder introduction sites -- params and
lam-params from the signature (param_tys), let-binders from the
LetBinderTypes table built in iter 1 -- via a new fn_modes_of helper
(strip_forall + Type::Fn { param_modes }). callee_arg_modes reads a
local Var callee's fn_param_modes before falling back to globals; the
existing App arg-walk maps Borrow to a borrow walk unchanged.
Scope call (orchestrator): all three seeding sites, not just the param
path that unblocks filter -- the point of this hardening cycle is to
leave no latent class (the #57-from-0063-deferral lesson), and the
extraction is the same at each site with the table already built.
Two tasks, RED-first: Task 1 adds examples/c1_local_hof.ail to the
harden list (RED: use-after-consume on filter_box's h); Task 2 adds the
field + seeding + callee_arg_modes read (GREEN), plus two in-source
unit tests (param path via signature, let path via hand-built table)
and a borrow-predicate test helper. Diagnostic-only; no schema/type
change. plan-recon mapped current post-iter-1 line numbers; parse gate
fired on the inlined fixture (RED confirmed).
refs #57
|
||
|
Brummel
|
47964abf7c |
feat(check): tee let-binder types, exempt value-typed let-binders (#57, 0064 iter 1)
First iteration of spec 0064 (the #55-cutover hardening, #57). Closes
false-positive class 3: a value-typed `let`-binder (e.g. a `Float`
result bound and read more than once) tripped `use-after-consume`
because the linearity walk installed every `let`-binder as a default
heap-tracked BinderState. Spec 0063 deferred exactly this class
("value-typed let-binders ... deferred until a corpus shape demands
it"); eqord_3_newton_sqrt.iterate's `(let xnew (app / …) …)` is that
shape.
Mechanism (the design call, user-delegated): the let-binder's type is
ALREADY computed at synth's Let arm and discarded. Stop discarding it.
synth gains a (def,binder)->Type out-param (LetBinderTypes, defined in
linearity.rs); the Let arm records `subst.apply(&v)`. The table is
allocated per-module in check_workspace, threaded through
check_in_workspace -> check_def -> check_fn/check_const/check_instance
-> synth, and passed (immutable) into check_module_with_visible ->
Checker. linearity's Term::Let seeds BinderState.is_value from the
table via the existing type_is_value predicate -- the same per-binder
flag #56 seeds at param/pattern/lam sites, now extended to the let
site. No schema change, no hash shift, no codegen change; the check
stays diagnostic-only.
Not a re-run of inference in the walk (ruled out by
|
||
|
Brummel
|
ba981689fc |
plan: 0064 iter 1 — let-binder type table + value-typed let exemption (#57)
First iteration of spec 0064 (#57 hardening). Tees the (def,binder)->Type table out of the typecheck pass (the inferred let-binder type computed at synth's Let arm and discarded today) and threads it into the linearity walk, then seeds BinderState.is_value for value-typed let-binders from the table -- closing false-positive class 3 (the class spec 0063 deferred). Two tasks, RED-first: Task 1 adds examples/c3_value_let.ail to the harden_ownership_false_positives_are_clean list (RED: use-after-consume on xnew); Task 2 threads the table end-to-end (synth -> check_fn -> check_in_workspace -> check_workspace -> check_module_with_visible -> Checker, all callers in one compile-atomic task) and seeds is_value at Term::Let (GREEN), plus two in-source unit tests (value-let clean, heap-let still errors). Fixes 1/2/4 (local fn-param modes, alias-redirect, partition_eithers rewrite) are later iterations of the same spec, explicitly out of scope here. plan-recon mapped the threading path; the parse-the-bytes gate fired on the inlined fixture (RED confirmed). refs #57 |
||
|
Brummel
|
f6a8607518 |
spec: 0064 harden ownership analysis part 2 (#57)
The #55 cutover (plan 0121, Task 4) is blocked: deleting
ParamMode::Implicit activates the strict linearity check universally,
but the migrated corpus does not pass cleanly. #56 (spec 0063) closed
two false-positive classes; universal activation surfaces three more
plus one genuine corpus over-consume. This spec is the "#56 part 2"
hardening, sibling to 0063, landing before the irreversible variant
deletion. It does NOT patch a fifth phase onto plan 0121 (project
rule: 2+ blocking classes = spec defect).
Four additive fixes, no schema/hash/codegen change, check stays
diagnostic-only:
- Type table teed from the typecheck pass (synth Let arm at lib.rs:3808
computes the binder type then discards it) into (def,binder)->Type,
threaded to linearity. The "stop discarding known information" fix,
not a re-run of inference in the walk (ruled out by
|
||
|
Brummel
|
39b674c1ec |
chore: throwaway mode-migration tool for the Implicit cutover (#55, 0121 task 3)
Adds `ail migrate-modes <file>` (a throwaway CLI subcommand) and the AST walker `ailang_core::ast::for_each_fn_type_mut` it drives: parse a .ail, map every bare/Implicit fn-type slot to Own, preserve explicit Own/Borrow, print back. Semantically invisible today (PartialEq treats Implicit == Own), but it materialises the modes so the post-cutover parser — which will reject bare slots — accepts the migrated corpus. Both are throwaway: removed in task 4 once ParamMode::Implicit is deleted (the closure references Implicit and would not compile). RED-first: migrate_modes.rs failed to compile (missing walker) before for_each_fn_type_mut was added. Additive; full workspace suite green. refs #55 |
||
|
Brummel
|
e1c908912b |
feat(check): reject borrow-returns at the signature (#55, 0121 task 1)
`(ret (borrow T))` is now a check error (CheckError::BorrowReturnNotPermitted, code `borrow-return-not-permitted`), fired on the fn signature before body dataflow. Borrow-returns are refcount-invisible and can outlive their source; re-enabling them needs the escape/liveness axis, which is out of scope (spec 0062). The diagnostic names that gap as an honest "not yet". The check_fn signature destructure now also binds `ret_mode` (param_modes stays under `..`; the borrow-over-value reject that reads it is folded into the cutover, task 4). RED-first: examples/borrow_return_reject.ail checked clean (exit 0) before the reject landed, exits 1 after. New CLI-boundary E2E pin crates/ail/tests/mode_signature_rejects.rs, modelled on raw_buf_new_type_arg_pin.rs. Additive — ParamMode::Implicit still present; full workspace suite green, bench/check.py 0 regressed. refs #55 |
||
|
Brummel
|
7dc21234f0 |
plan: re-stage 0121 — fold borrow-over-value into the cutover (mono collision)
First implementation contact surfaced a plan defect: the
borrow-over-value reject cannot land green as a standalone pre-cutover
task. The prelude's polymorphic comparison builtins are
`(params (borrow a) …)` (eq/compare/lt/…); the mono pass
(apply_subst_to_type, subst.rs:165) specialises `a := Int/Bool/Float`
into `compare__Int(borrow Int, …)` — the language's OWN generated code
in the forbidden `(borrow value-type)` shape. The standalone check
broke compare_{int,bool}_mono_symbol_emits_branch_ladder.
Spec 0062's "value types always own" rule is universal, so generated
code must honour it too: the principled fix is a mono-pass coercion
`(borrow value-type) -> (own value-type)` (a no-op — value types carry
no RC, so own vs borrow emits no inc/dec and the branch-ladder IR is
unchanged). The check and the coercion are a matched pair and now live
together in Task 4 (Steps B1-B2), which touches the mono path anyway.
Pre-cutover green tasks are now Task 1 (borrow-return reject) + Task 3
(throwaway migration tool). Task 1's check_def destructure binds only
ret_mode (param_modes stays under `..`); Task 4 Step B2 widens it.
Spec gap to record at the next 0062 brainstorm touch: the
borrow-over-value rule needs its mono-coercion clause spelled out — the
spec tested only the authored case.
refs #55
|
||
|
Brummel
|
a84ba596cf |
plan: eliminate the Implicit ownership default (0121)
Executable projection of spec 0062 (#55) into four tasks: (1) the borrow-return reject, (2) the borrow-over-value reject, (3) a throwaway `ail migrate-modes` pass (parse -> Implicit->Own, keep explicit own/borrow -> print explicit), (4) the atomic cutover (run the migration over the corpus + hand-fix the read-only-heap params the now-universal linearity check flags, delete the variant + all compile sites compiler-driven, parser bare-slot reject, reset the hash pins, flip the leak pin, update both contracts, drop the throwaway). Placeholder-free; all seven surface fixtures parse-gated against HEAD. Spec 0062 marked approved (user, 2026-06-01) and its soundness re-validated against post-#56 HEAD: the three own-return-provenance / regime-A fixtures still exit 1 under [consume-while-borrowed]; #56's application-is-a-borrow change does not weaken them (none is an application of a borrowed binder). plan-recon folded four spec-enumeration corrections into the plan: - kernel migration target is raw_buf/source.ail, not the retired kernel_stub/source.ail the spec cited; - FIVE hash pins shift, not the two the spec named (embed_export_hash, eq_ord_e2e, mono_hash_stability are the missed twins); - design/contracts/0002-data-model.md also documents the "implicit" JSON form and is drift-anchored -> updates alongside 0008; - real counts are 261 fn-type fixtures (spec ~208) and 17 fn_implicit references (spec 16); the compiler is the authoritative enumerator. refs #55 |
||
|
Brummel
|
aaa70d4c35 |
feat(check): harden the ownership analysis for universal activation (0063)
The strict linearity check (use-after-consume / consume-while-borrowed, linearity.rs) runs today only on the ~45 of 258 all-explicit-mode fns; its activation gate skips any fn with a bare/Implicit param. Deleting ParamMode::Implicit (#55) would turn it on universally and surface ~21% false positives in two classes. This closes both, making the core ownership analysis sharp across the whole codebase for the first time — the precondition for #55. Diagnostic-only: no schema, no hash, no codegen change; the two existing diagnostic codes simply fire on a smaller, more correct set. Fix 1 — value-type exemption. A new is_value_type predicate (ailang-core::primitives) names the unboxed set {Int,Bool,Float,Unit}; BinderState gains an is_value flag seeded from the binder's locally available type (param signatures, ctor field types for pattern binders, lam typed-params), and use_var short-circuits the Consume arm for it. A value type has no refcount and is never consumed, so multi-read is always legal. Str is deliberately NOT in the value set, though is_primitive_name includes it: drop.rs:490-492 lowers only Int/Bool/Float/Unit to non-ptr; Str is a ptr, RC-dec'd, so a multi-consume of Str without clone is a real use-after-free. is_value_type is the Str-excluding subset of is_primitive_name, pinned by a unit test. is_heap_type and the over-strict-mode lint are left untouched (separate concern). Fix 2 — application is a borrow. Term::App walks its callee in Position::Borrow (was Consume). Applying a function value reads it; it stays live for further applications. Global fn-refs are untracked (no-op); a tracked function-typed binder (a HOF param) is no longer consumed by application. This fixes both the own-f-param use-after-consume and the borrow-f-param consume-while-borrowed in the recursion-passing HOF shape (map_int f t). Passing a function value as an arg is unchanged — it follows the callee param_modes. Scope decision: value-typed let-binders stay conservatively tracked as heap (their type is inferred, not annotated, and the linearity walk does not re-run inference). This is soundness-safe — over-strict, never unsound — and not a measured corpus shape; lifting it would need a full binder->type table out of the type-checker. Verification: all three false-positive classes reproduced today against explicit-mode fns and shipped as RED->GREEN fixtures (examples/fp_{value,hof,map}.ail); examples/real_consume.ail stays RED (heap double-consume still fires) to prove the exemption is type-gated. 715 workspace tests green; bench/check.py 0/34 and bench/compile_check.py 0/24 regressed. merge_states carries is_value so the flag survives branch merges. RED-first verified per task. closes #56 |
||
|
Brummel
|
47fb328aca |
plan: harden ownership analysis — is_value_type + value exemption + app-borrow (0120)
Executable projection of spec 0063 into four tasks: (1) is_value_type predicate in ailang-core::primitives; (2) application-is-a-borrow (Term::App callee walked Position::Borrow); (3) value-type exemption (BinderState.is_value seeded at param/pattern/lam sites, use_var short-circuit); (4) on-disk RED->GREEN fixtures + workspace assertions. Placeholder-free with exact code for every edit site. refs #56 |
||
|
Brummel
|
8cdac7ecef |
spec: harden the ownership analysis for universal activation (0063)
Precondition for #55 (eliminate ParamMode::Implicit, spec 0062). The strict linearity check (use-after-consume / consume-while-borrowed) runs today only on the ~45 of 258 all-explicit-mode fns; deleting Implicit turns it on universally and surfaces ~21% false positives in two classes — value-type params (Int/Bool/Float/Unit, never consumed) and applied function params in HOFs (application is a borrow). Two design questions settled against the live tool, not the model: - Str is heap, NOT a value type for consume-tracking. drop.rs:490-492 lowers only Int/Bool/Float/Unit to non-ptr; Str is a ptr, RC-dec'd, so a multi-consume of Str without clone is a real use-after-free. The value-type set is the unboxed {Int,Bool,Float,Unit}, narrower than is_primitive_name (which includes Str). - Applying a function value is a borrow; passing it as an arg follows callee param_modes (unchanged). The recursion-passing HOF pattern is then consistent under a borrow f-param. All three false-positive classes reproduced today against explicit-mode fns (testable without #55) and recorded as RED fixtures; a heap double-consume fixture stays RED to prove the exemption is type-gated. grounding-check PASS. refs #56 |
||
|
Brummel
|
6c351af169 |
docs(claude): record the LLM-designed mandate and last-resort escalation
AILang is built primarily for LLMs; the corollary the owner stated is that it is also meant to be designed by them. The human owner is escalated to only as a last resort, and at that point the fix is most likely a revert rather than a redesign. Records the design-authority half of the identity that the opening section already stated for authoring. |
||
|
Brummel
|
c8b765614d |
spec: eliminate the Implicit ownership default (0062)
Design spec for #55, superseding #54. Deletes ParamMode::Implicit in a single cutover -> binary {Own, Borrow}; no defaulted ownership mode survives on any fn-type slot, authored or compiler-synthesised. The grounding pass corrected the central architecture claim. Both #55 and the first draft assumed a new return-provenance check was needed to reject an own-return aliasing a borrowed binder, "silently accepted today". Running the candidate fixtures through the live `ail check` falsified that: consume-while-borrowed already rejects all three provenance paths -- direct passthrough, let-indirection, and borrow escaping into a returned constructor (regime A). Own-return soundness and regime A are already-green properties. The cutover therefore adds exactly one new check (borrow-return rejection) plus one signature-level reject (borrow-over-value), not two new provenance analyses. Scope decisions ratified: keyword stays verbal (own/borrow), ParamMode keeps its name; totality is strict including value-type slots (every slot moded, (own value) trivial, (borrow value) an error, bare slot a parse error) so an LLM author needs no heap/value knowledge. The leak fix (spec 0061 symptom) falls out of Implicit->Own at the synthesis sites: the old typechecker made Implicit and Own indistinguishable (mode_eq), so setting Own changes no typecheck outcome -- it only flips the codegen dec gate from skip to emit. Out of scope: re-enabling borrow-returns (needs the escape/liveness axis), multi-return, mode polymorphism. grounding-check: PASS (7 assumptions ratified against named green tests; all fenced ail blocks validated against the live tool). refs #55 |
||
|
Brummel
|
657b24bfc9 |
docs(design): refine the ownership-totality model from design discussion (0008)
A design discussion on 0008 surfaced four refinements; this records them in the exploratory model. Still design exploration, not a spec commitment. - §3.4 (new): the keywords are adjectival, not verbal. `own`/`borrow` are acquisition verbs — native on a parameter (the callee acquires), backwards on a return (the callee disposes; §3.1's own text switches feet between the two readings). Naming the mode adjectivally (`owned`/`borrowed`) describes the value the receiver holds and is position-independent, strengthening §3's uniformity rather than breaking it. Carries a safety stake, not only readability: a wrong `own` is the double-free (§6), and the verb form is a mis-annotation pump pointed that way. Notes the corroborating `ret_mode: ParamMode` type-name smell (a return mode is a covariant promise, not a contravariant param demand). - §4.4: "fixed modes suffice" is contingent on the point-free cut, not a standalone theorem. Adds the structural discriminator — today's HOFs control both ends of their seam, so the seam mode is determinate; only the wire-two-black-boxes kind (point-free) needs a mode variable. - §5.1–5.3 (new): a tuple question exposed that `ret_mode` is flat but ownership is a tree over the type — a single top-level `Own` on a mixed Pair (owned field + borrowed field) decs borrowed data on drop. Resolved to regime A (borrows may not escape into an escaping heap structure → transitive ownership → flat `ret_mode` sound), not a per-field mode tree (regime B = lifetimes). For compound returns, verification is constitutive of the flat representation's meaning, not a separable phase — sharpening Q2 to "must" there. Multi-return is the ownership-honest return path (flat per-position modes, no reachability pass) but does not subsume the typed boxed aggregate; the boxing point is the ownership commit. - §7: Q2 sharpened (compound returns force it), Q3 rewritten (adjectival keywords + the pre-cutover timing, since a post-cutover rename is a second irreversible hash reset), Q4 added (multi-return adoption); resolved-footer extended with regime A and multi-return≠tuple. - §8 (new): relation to Rust. Concedes the primitive-level convergence (own/borrow ≈ move/borrow, regime A ≈ Rust's escape rule, unboxed product ≈ tuple, RC ≈ Rc/Arc), then names what AILang drops and why: no lifetime polymorphism (the point-free cut removes the pressure that makes Rust carry it), no `&mut`/shared-XOR-mutable (no shared mutable refs; mutation rides uniqueness, the Clean lineage), no elision (§2's no-defaults is the opposite of Rust's elision). Scoped to the totality-relevant divergences; the broader memory-model lineage is cross-referenced to 0004 rather than duplicated. |
||
|
Brummel
|
768189abce |
docs(design): add the ownership-totality whitepaper (model 0008)
Exploratory model that converges the direction sparked by the
typed-MIR fieldtest Implicit-return leak (spec 0061). It is not
current state and not an approved milestone — the surface still
defaults ParamMode::Implicit; the file fixes a future shape before
committing it to specs.
The converged shape:
- Ownership is authored, never defaulted. Implicit is removed from
the authorable surface; every parameter and return carries
own/borrow. A default cannot be half-removed, so totality is the
only coherent stopping point.
- Value types take trivial-own; borrow over a value type is an
error. The surface stays uniform so the author needs no knowledge
of which types are heap.
- No mode-polymorphism. A type-variable position is never a hole
under fixed modes — it takes a concrete own/borrow. Mode variables
would only remove duplication, and the reuse-across-modes need is
confined to point-free combinators (cut by the language identity)
and the consume-vs-preserve container duality (better expressed as
two named functions). Structurally, AILang iterates via tail-app,
not passed-function combinators, so the higher-order own/borrow
clash does not arise. ParamMode stays binary {Own, Borrow}.
Left open: cutover shape, verification sequencing, glossary naming.
The one irreversible consequence — a corpus-wide module-hash reset
once Implicit stops eliding the mode fields from canonical JSON — is
named for deliberate sign-off.
|
||
|
Brummel
|
7a99c5dbfa |
docs(fieldtest): record the typed-MIR milestone-close fieldtest
Milestone-close fieldtest of the typed-MIR work, run by a downstream LLM author exercising only the public interface. Four curated scenarios plus a reduction-bisect set, derived top-down from the milestone promise (heap-Str-across-recur, new-over-user-ADT, cross-module print__<UserType>, combined render loop). Three [working] findings ratified: new-over-user-ADT + Show + print (#51/#53 class), cross-module synthesised print, and heap-Str-across-recur value correctness all produced correct output first try. One [bug] found and orchestrator-verified: returning an owned heap-Str from a callee fn across the call boundary leaks exactly one RC slab (live=1) — producer-independent (str_concat, int_to_str), loop-independent, silent (output correct). It sits in the coverage gap the curated *_no_leak_pin fixtures leave open (they keep the heap-Str expression inside main). Distinct from #49 (loop-carried within one frame); this is the function-return leg. Routed to debug RED-first; blocks the deliberate milestone-close until GREEN. |
||
|
Brummel
|
bd62f2b9d0 |
docs(design): name the fn-value resolution carve-out in 0018
Audit follow-up (typed-MIR milestone-close drift review, clean). The boundary contract said codegen does "no callee resolution"; that is true for App callees — the only thing `Callee` totality claims over — but `resolve_top_level_fn` still resolves a dotted fn used as a *value* via `import_map` (the in-corpus-unreachable type-home residue the spec blessed at mir.2), and ctor/const resolution stay codegen-side. Tighten the prose to "no App-callee resolution" and name the value-position carve-out, so the contract is precise and a future drift review does not read the blanket claim as violated. Prose-only; no code, no test change. |
||
|
Brummel
|
a378dad0aa |
docs(design): ratify the check→codegen boundary (mir.5, typed-MIR close)
mir.5 is the typed-MIR milestone's closing iteration. Its CODE half had
already converged before the iteration began, so mir.5 ships no code —
it ratifies into the design/ ledger the boundary the code already holds.
Verified at iteration entry (empirically, not from the spec sketch):
- all four named re-derivers grep-clean in codegen: synth_with_extras,
synth_arg_type, type_home_module, the second infer_module_with_cross;
- lower_workspace takes &MirWorkspace (codegen consumes MIR);
- MTerm::New is unreachable!() — raw-buf.4 desugars Term::New to
(app T.new …) before codegen, so there is no element-type
re-derivation left to relocate;
- #51 / #53 (the element-type / new-T codegen crashes) are closed;
their residue was fixed by
|
||
|
Brummel
|
c5fd16a4eb |
feat(codegen): StrRep loop-carried Str ⇒ Heap, close the #49 recur leak
mir.4 closes bug #49 (a heap-Str loop binder replaced across `recur`
leaks every superseded slab; the loop result leaks too) structurally,
by making every Str a loop binder holds OR a loop returns an owned heap
slab — then deleting the `!is_str` carve-out from BOTH codegen gates
that carried it. The leg-by-leg `!is_str` patches are gone.
Mechanism (the milestone thesis: representation in MIR, codegen reads
it). lower_to_mir sets rep = StrRep::Heap on every Str literal in a
loop-carried position; codegen's MTerm::Str arm promotes a Heap literal
via ailang_str_clone (a fresh ailang_rc_alloc slab, rc_header refcount
1). Three promotion positions, all in the Loop/Recur lowering:
1. binder seed inits (is_str_ty on the binder type);
2. recur args at Str-binder positions (read off the innermost
loop_stack frame);
3. exit-arm tail result literals (promote_tail_str_literals walks the
loop body's tail positions when the loop returns Str).
With every loop-carried Str now owned-heap, both gates lose !is_str:
the recur superseded-value dec (lib.rs) frees each intermediate slab;
the loop-result trackability gate (drop.rs:493) lets the caller's
scope-close free the final slab.
Why both gates, and the planning-time error this corrects. The original
plan/spec scoped the deletion to the recur dec gate only, reasoning the
#49 loop result is "consumed by print". The implement-orchestrator
landed that scope (Tasks 1-2) clean and correctly BLOCKED: it took #49
from live=3 to live=1, the residual being the loop RESULT. I verified
the diagnosis empirically:
- print BORROWS its arg (prelude: `(let s (show x) (do io/print_str
s))`, the eob.1 heap-Str discipline), so `(print s)` does not free
the loop result; the caller's let-binder does, via drop.rs:493.
"consumed by print" != "freed by print" — that was the error.
- Deleting drop.rs:493's !is_str takes #49 and the recur-literal
fixture to live=0.
- Deleting drop.rs:493 WITHOUT exit-arm promotion SIGSEGVs (exit 139)
on a loop returning a static Str literal — hence the third
promotion position and the static-exit witness.
The agent surfaced a real spec defect via an empirical BLOCK rather
than guessing; I corrected the spec refinement note (both gates, three
positions) and completed the loop-result leg inline, the context
already loaded from verifying the BLOCK (CLAUDE.md "already loaded
context" carve-out). drop.rs:493's Str carve-out is now sound to delete.
Boundary (asserted ill-typed, not constructible): a borrowed static-Str
Var seeded/recur'd into a consumed Str loop binder — rejected upstream
by uniqueness (a consumed binder position is owned).
Verification (orchestrator-run): all four leak fixtures reach live=0
under AILANG_RC_STATS — #49 (xyyy), recur-literal (reset), static-exit
(result), and the
|
||
|
Brummel
|
2536b5f535 |
plan(mir.4): StrRep loop-carried Str ⇒ Heap, delete the recur !is_str gate
mir.4 closes bug #49 (a heap-Str loop binder replaced across `recur` leaks every superseded str_concat slab, live=3) structurally rather than with a fourth leg-by-leg patch. Design calls made in planning (folded into the spec as the mir.4 refinement note): - Mechanism: producer-side representation, codegen reads it — the milestone thesis. lower_to_mir sets rep = StrRep::Heap on every MTerm::Str literal that flows into a Str loop-binder alloca (seed inits AND recur args at Str-binder positions); codegen's MTerm::Str arm gains a Heap leg that promotes the static literal via ailang_str_clone (fresh ailang_rc_alloc slab, rc_header refcount 1). No special-casing in the codegen Loop/Recur store arms — the clone is emitted automatically when the Str{Heap} node is lowered. - Recur args are promoted too, not just seeds: `(recur "reset" …)` is well-typed (verified: `ail check` accepts it), and a seed-only promotion would leave a static literal under the now-unconditional dec — a constructible UB hole. Soundness over minimalism on the highest-risk RC/drop surface. - Only the recur dec gate (lib.rs:2231) loses !is_str. The drop.rs loop-RESULT trackability gate (drop.rs:493) STAYS conservative: deleting it needs a broader "every Str a loop can return is owned-heap" guarantee (a static exit-arm literal breaks it) and is not required by the #49 acceptance (the #49 loop result is consumed by print). drop.rs is untouched. - Boundary (out of scope, asserted ill-typed): a borrowed static-Str Var seeded/recur'd into a consumed Str loop binder is rejected upstream by uniqueness (a consumed binder position is owned), so it is not constructible. Plan 0119 carries exact code per step: producer rep promotion + a loop-frame-driven Str-binder mask for recur args (reusing the existing ctx.loop_stack, no new ctx field), the codegen Heap leg, the gate deletion, the #[ignore] lift on the #49 pin, a recur-literal witness fixture + runtime pin proving the recur-arg leg, and the full-suite + ir_snapshot verification. The existing lower_to_mir_ty Static-seed pin flips to assert Heap. |
||
|
Brummel
|
eba1be8d9d |
feat(codegen): fill MArg.mode for App args, read the anon-temp drop gate off it, delete module_def_ail_types (mir.3b)
Second of two iterations delivering spec 0060's mir.3 row (plan docs/plans/0118-mir.3b-marg-mode-and-table-delete.md). lower_to_mir's Term::App arm fills each MArg.mode from the resolved callee's param_modes (the sig = synth_pure(callee) it already holds); codegen's emit_call anon-temp borrow-slot drop gate reads arg.mode instead of re-looking-up the callee's param_modes from the module_def_ail_types table; and — since that gate was the table's only reader — the module_def_ail_types field plus all its construction and threading are deleted. Two refinements to the spec's mir.3b sketch, settled in planning from a focused recon and recorded in spec 0060: 1. module_def_ail_types is deleted in mir.3b, not mir.5. A grep proved self.module_def_ail_types had exactly ONE reader (the anon-temp gate); the own-param drop reads param_modes from the def's own f.ty, not this table. Once the gate moves onto MArg.mode the table is dead, so it is removed now rather than carried as dead code to mir.5. The mir.5 row's "last re-derivation residue" shrinks to the element-type / Term::New (New.elem) work. 2. MArg.mode is filled for App args only; MTerm::Let.mode is not filled. Only App args have a real per-arg mode source (the callee Type::Fn.param_modes). Do args (EffectOpSig has no param modes), Ctor args (no per-field mode), and Recur args (loop binders carry no mode) stay Mode::Owned. Let.mode has no source (ast::Term::Let has no mode field) and no consumer (the let-drop gate reads consume, not Let.mode), so it stays Owned — filling it would invent a value nobody reads. Safety property held: MArg.mode is filled from the same callee fn-type the gate read from the table, so the drop fires identically. For (app RawBuf.get (app RawBuf.set …) 0), RawBuf.get's param 0 is borrow, so args[0].mode = Borrow — exactly the value module_def_ail_types yielded. Confirmed by the anon-temp witness staying green. ParamMode -> Mode conversion: Borrow -> Mode::Borrow; Own and Implicit (the Implicit ≡ Own contract) -> Mode::Owned. The own-param drop keeps reading ParamMode off f.ty (Type/ParamMode imports retained, live readers at lib.rs:1356/1507). Also retires three now-stale doc comments that named the deleted module_def_ail_types field (ailang-check/src/lib.rs, uniqueness.rs, and the codegen_import_map_fallback_pin test doc) — a direct consequence of the deletion, reworded to the current architecture. Verification (orchestrator, post-implement inspect): diff matches the plan (App-arg m_args built before m_callee consumes sig — the benign borrow-order deviation the plan's note flagged); module_def_ail_types grep-clean across all of crates/; cargo build --workspace clean (no unused/dead_code); cargo test --workspace 702 passed / 0 failed / 3 ignored (+1 = the new producer pin app_arg_carries_callee_borrow_mode; no #[ignore] added). The anon-temp drop-correctness witness raw_buf_owned_drop_balances_rc_stats stays live=0, now driven by MArg.mode; the other RC-stats leak pins green; #49 stays #[ignore] (mir.4); #51/#53 build guards green. mir.3 is now complete (3a relocated consume_count + deleted the second uniqueness run; 3b filled the mode annotation + deleted module_def_ail_types). Remaining: mir.4 (#49 StrRep RED->GREEN), mir.5 (element-type/Term::New + ledger). |
||
|
Brummel
|
0bd7f47fad |
plan: mir.3b fill MArg.mode, switch the anon-temp gate onto it, delete module_def_ail_types
Executable plan for the second of two iterations delivering spec 0060's mir.3 row. A focused recon (the anon-temp gate + the per-arg mode sources) established two refinements to the spec sketch, both locked in the plan: 1. module_def_ail_types is deleted in mir.3b, not mir.5. A grep proved self.module_def_ail_types has exactly ONE reader — the emit_call anon-temp borrow-slot drop gate. The own-param drop reads param_modes from the def's own f.ty, not this table. So switching that gate onto MArg.mode leaves the table dead; deleting it now is the clean move. mir.5's "last re-derivation residue" shrinks to element-type/Term::New. 2. MArg.mode is filled for App args only; MTerm::Let.mode is not filled. Only App args have a real per-arg mode source (the callee Type::Fn.param_modes). Do/Ctor/Recur args have no per-arg mode (EffectOpSig/Ctor.fields/loop-binders carry none) -> stay Owned default. Let.mode has no source (ast::Term::Let has no mode field) and no consumer (the let-drop gate reads consume, not Let.mode) -> stays Owned. Safety property (recon-confirmed on the #43 anon-temp witness): MArg.mode is filled from the same callee fn-type the gate read from the table, so the drop fires identically. For (app RawBuf.get (app RawBuf.set …) 0), RawBuf.get's param 0 is borrow -> args[0].mode = Borrow, exactly the value module_def_ail_types yielded. Four tasks: (1) producer fills App-arg MArg.mode via a param_mode_at helper; (2) codegen gate reads arg.mode + delete module_def_ail_types (field + construction + threading, compiler-completeness-checked); (3) producer pin (App arg mode == Borrow) + the live=0 anon-temp acceptance (raw_buf_owned_drop_balances_rc_stats); (4) record the refinements in spec 0060. No new .ail fixture (reuses raw_buf_drop_min.ail). |
||
|
Brummel
|
6bc0b501cb |
feat(codegen): relocate per-binder consume_count into MIR; delete codegen's second uniqueness run (mir.3a)
First of two iterations delivering spec 0060's mir.3 row (plan docs/plans/0117-mir.3a-consume-count-relocation.md). The per-binder consume_count — the only thing codegen read from its second infer_module_with_cross run — now lives on a new binder-keyed MirDef.consume: BTreeMap<String,u32>, filled once by lower_to_mir on the post-mono body. Codegen builds its existing (def,binder)->consume_count lookup from those maps and the three scope-close drop sites (own-param, let-binder, match-arm) read it; the codegen uniqueness run, its UniquenessTable field, and the import are deleted. mir.3 is split into two iterations (recorded in spec 0060): the named deletion depends ONLY on consume_count. All three drop sites read consume_count from self.uniqueness, while the modes they also gate on (param_modes for the own-param dec, scrutinee_is_owned for the match dec) come from the type, not the uniqueness pass (UniquenessInfo carries no mode). So mir.3a relocates consume_count and ships the deletion; mir.3b (next) fills MArg.mode/MTerm::Let.mode and switches emit_call's anon-temp gate onto MArg.mode. The split halves the change at the codebase's highest-risk surface (RC/drop correctness, the raw-buf leak saga) and makes each half independently verifiable against the live=0 leak pins. Design: consume_count lands on MirDef (binder-keyed), NOT on MArg. The drop sites key by (def, binder_name) — a per-binder aggregate — while the spec-sketched MArg.consume_count is per-arg-position and never reaches them; the per-def map matches the UniquenessTable's own keying and all three binder kinds (param/let/pattern) uniformly. MArg.consume_count stays a mir.1 default. Source = run check's infer_module_with_cross inside lower_to_mir post-mono (the synth_pure single-engine-re-walked pattern); retaining a check-time result is blocked (check's uniqueness is pre-mono and runs on-demand-from-codegen, never in check_workspace, so it would miss the mono specialisations). cross_module_types (= codegen's module_def_ail_types) is rebuilt in elaborate_workspace from the post-mono workspace. Safety property held: this is a PURE RELOCATION of an identical computation — infer_module_with_cross ran on the post-mono module in codegen and now runs on the same post-mono module in lower_to_mir, so drop placement is byte-identical. cross_module_types matches module_def_ail_types exactly (both insert f.name->f.ty over post-mono Def::Fn), confirmed by the leak pins staying green. Verification (orchestrator, post-implement inspect): diff matches the plan; codegen grep-clean for infer_module_with_cross + UniquenessTable (now ailang-check-internal); cargo build --workspace clean (no errors, no unused warnings — module_def_ail_types survives for emit_call's anon-temp param_modes gate); cargo test --workspace 701 passed / 0 failed / 3 ignored (+1 = the new producer pin mirdef_consume_is_populated_for_let_binder; no #[ignore] added). The live=0 drop-correctness witnesses all green: alloc_rc_loop_valued_rawbuf_binder_drops_at_scope_close (#43/#47), alloc_rc_heap_loop_binder_replaced_by_recur_does_not_leak, alloc_rc_print_int_does_not_leak_show_result_str. #49 stays #[ignore] (mir.4); #51/#53 build guards green. |
||
|
Brummel
|
69749fc3e1 |
plan: mir.3a relocate consume_count into MIR, delete codegen's second uniqueness run
Executable plan for the first of two iterations delivering spec 0060's mir.3 row. Two plan-recon passes (mode/consume/drop-placement surface + the uniqueness-pass signature) established that the named deletion (the second infer_module_with_cross run) depends ONLY on consume_count: all three codegen scope-close drop sites read consume_count from self.uniqueness, while the modes they also gate on (param_modes for the own-param dec, scrutinee_is_owned for the match-arm dec) come from the type, not the uniqueness pass (UniquenessInfo carries no mode). So mir.3 splits: - mir.3a (this plan): relocate consume_count; delete the second run. - mir.3b (next): fill MArg.mode / MTerm::Let.mode from the type modes and switch emit_call's anon-temp borrow-slot gate onto MArg.mode. The split halves the change at the codebase's highest-risk surface (RC/drop correctness, where the raw-buf leak saga lived) and makes each half independently verifiable against the live=0 leak pins. Design decisions locked in the plan: - consume_count lands on a new binder-keyed MirDef.consume: BTreeMap<String,u32>, NOT on MArg. The recon surfaced the central structural tension: the drop sites key consume_count by (def, binder_name) — a per-binder aggregate — while the spec-sketched MArg.consume_count is per-arg-position and never reaches them. The per-def binder map matches both the UniquenessTable's own keying and all three binder kinds (param/let/pattern) uniformly. MArg.consume_count stays a mir.1 default (no consumer). Spec 0060 is updated to record this (Task 5). - Source = run check's infer_module_with_cross inside lower_to_mir on the post-mono module (the synth_pure single-engine-re-walked pattern). The alternative of retaining a check-time result is blocked: check's uniqueness runs pre-mono and on-demand-from-codegen, never during check_workspace, so a retained result would miss the mono specialisations. cross_module_types (= codegen's module_def_ail_types, module->def->Type) is rebuilt in elaborate_workspace from the post-mono workspace. Key safety property carried in the plan: this is a PURE RELOCATION of an identical computation. infer_module_with_cross ran on the post-mono module in codegen and now runs on the same post-mono module in lower_to_mir — same function, same input, same output — so drop placement is byte-identical. The live=0 leak pins (raw_buf_loop_no_leak, loop_recur_heap_binder, print_no_leak) are the value-correctness gate. Five tasks: (1) add MirDef.consume; (2) elaborate_workspace builds cross_module_types + lower_module runs the pass and fills consume; (3) codegen builds its (def,binder)->consume lookup from MIR, the three drop sites read it, delete the run + field + import; (4) producer pin + leak-pin acceptance + grep-clean; (5) record the split in spec 0060. No new .ail fixture (producer pin reuses new_counter_user_adt.ail). |
||
|
Brummel
|
a6fd93adba |
feat(codegen): pre-resolve the App callee in MIR; delete is_static_callee + type_home_module (mir.2)
Second re-deriver removal of the typed-MIR milestone (spec
docs/specs/0060-typed-mir.md, plan docs/plans/0116-mir.2-callee-static.md).
Static-callee resolution moves out of codegen and into lower_to_mir:
every Term::App callee is classified once, mirroring check's own synth
Term::Var ladder (lib.rs:3409-3582), and emitted as a resolved Callee.
Codegen reads the callee identity off MIR and routes each kind to the
right lowering. The codegen re-derivers is_static_callee and
type_home_module are deleted, and the lower_app <-> is_static_callee
lockstep pair is struck from CLAUDE.md.
Callee reshape (beyond the spec's original {module, fn_name} sketch —
spec 0060's Concrete-code-shapes block is updated in this iteration):
pub enum Callee {
Static { module, fn_name, sig: Type }, // user/prelude fn -> emit_call, module pre-resolved
Builtin { name, sig: Type }, // operator/intrinsic -> inline opcode lowering
Indirect(Box<MTerm>), // dynamic: fn-pointer / closure / shadowed name
}
Two forced refinements, both settled in planning:
- sig:Type on each resolved variant. drop::synth_callee_ret_mode reads
the callee ret_mode, today off Callee::Indirect(inner).ty(). A
resolved callee has no sub-term, so the sig (= synth_pure(callee),
mode-preserving) is the lossless replacement. This is NOT a mir.3
pull-forward: the MArg param-mode / consume_count fields stay at their
mir.1 defaults. Without it, every statically-resolved call would lose
its drop signal and the RawBuf / int_to_str RC-stats pins mir.1b fixed
would regress.
- Builtin variant. Operators and the str-num builtins (+, not,
str_concat, ...) are lowered inline by name and have no module/fn_name;
a separate variant is cleaner than a sentinel module. The classifier
decides Builtin vs Static from CHECK'S OWN resolution (a name in
env.globals with no owning module is a builtin), never a copy of
codegen's old is_static_callee allowlist. The recon's divergence audit
confirmed check's builtin set and codegen's inline-arm set match
exactly, so the single-engine rule is mechanically satisfiable with no
env threading gap.
type_home_module is fully deletable: a workspace-wide grep confirmed no
program references a type-scoped T.fn as a value, so its only other
consumer (resolve_top_level_fn) collapses to the module-name fallback
with no behaviour change. (The spec's "x3 mirrors" wording over-counted;
the live surface was the definition plus two call sites.)
Alternatives rejected during planning: sentinel module in Static
(ugly); builtins left as Indirect (breaks — no fn-pointer for an
operator); name-rewrite Static{name,sig} keeping lower_app's by-name
dispatch (blurs builtin/user and still needs sig). The mir.1b
re-synth-strictness trap (post-mono synth_pure re-unify surfacing
mode-strip / bare-vs-qualified / metavar leaks) did NOT fire under the
new producer path — qualify_local_types mode-preservation and the $u
wildcard normalisation from mir.1b held.
Verification (orchestrator, post-implement inspect): diff matches the
plan; grep-clean for is_static_callee + type_home_module across
ailang-codegen and ail; cargo build --workspace green; cargo test
--workspace 700 passed / 0 failed / 3 ignored (no #[ignore] added this
iteration — the 3 are the pre-existing #49 Str-leg pin + two doctests).
Acceptance witnesses green: #53 (mono_new_over_user_adt_builds_and_runs),
#51 (rawbuf_new_int_size_only_builds_and_runs still builds), show_print
cross-module (print_user_adt_runs_end_to_end), the new mir.2 pins
(mir2_resolved_callee_kinds_run_end_to_end, callee_classification_*,
strengthened new_over_user_adt_carries_node_types asserting Callee::Static).
#49 heap-Str loop binder stays #[ignore] (lifts at mir.4).
Two new examples/ fixtures back the new pins (classify_pin.ail for the
producer classification pin, mir2_callee_kinds.ail for the E2E),
ail-parse / round-trip clean.
Forward note (cycle-close architect): crates/ail/tests/codegen_import_map_fallback_pin.rs:14-15
doc prose names lower_app's cross-module arm and synth_with_extras as
failure-mode targets — both now deleted (mir.2 / mir.1b). The pin's
protected invariant is still valid and green; the prose is stale and
wants a doc-honesty reword, left out of this iteration's footprint.
|
||
|
Brummel
|
d81ea93de6 |
plan: mir.2 Callee::Static pre-resolved, codegen stops re-deriving the callee
Executable plan for spec 0060's mir.2 iteration: relocate static-callee
resolution from codegen into lower_to_mir. Two plan-recon passes (codegen
consumer surface + check-side producer surface) mapped the resolution
frontend; this plan locks three design decisions made during planning,
ahead of the tasks:
1. The Callee bridge enum is reshaped beyond the spec sketch: a third
variant Builtin{name,sig} (operators / str-num intrinsics have no
module/fn_name and are lowered inline by name), and a sig:Type on each
resolved variant. The sig is the lossless replacement for the pre-mir.2
Callee::Indirect(inner).ty() read that drop's synth_callee_ret_mode
depends on; it is synth_pure(callee), mode-preserving. NOT a mir.3
pull-forward — the MArg param-modes stay at mir.1 defaults. Spec
0060's Concrete-code-shapes Callee block is updated as part of the
iteration (Task 5).
2. Classification mirrors check's own synth Term::Var ladder
(lib.rs:3409-3582), never copies codegen's is_static_callee allowlist.
The recon's divergence audit confirmed check's builtin set and
codegen's inline-arm set match exactly, so the single-engine rule is
mechanically satisfiable with no env threading gap.
3. type_home_module is fully deletable: a workspace-wide grep confirmed no
program references a type-scoped T.fn as a value, so its second
consumer (resolve_top_level_fn) collapses to the module-name fallback
with no behaviour change. The spec's "grep-clean" acceptance holds; the
spec's "x3 mirrors" wording over-counts (def + 2 live sites).
Five tasks: (1) reshape Callee; (2) lower_to_mir classify_callee +
App arm; (3) the atomic codegen consumer switch (App arm/drop/lower_app
split/delete is_static_callee+type_home_module/resolve_top_level_fn);
(4) pins (strengthen #53 to assert Callee::Static, add a three-way
classification pin) + workspace suite; (5) strike the lower_app <->
is_static_callee lockstep row from CLAUDE.md, update spec 0060, clean
stale comment references. The classify_pin fixture is ail-parse-gated
(exit 0).
|
||
|
Brummel
|
895ba846e8 |
feat(codegen): switch the lowering walk from &Term to typed &MTerm (mir.1b)
Atomic completion of spec iteration mir.1 (docs/specs/0060-typed-mir.md): codegen now consumes the typed MIR produced by `lower_to_mir` instead of re-deriving types from the bare `ast::Term`. Every codegen helper that took `&Term` (lower_term, lower_app, drop.rs, match_lower.rs) takes `&MTerm` and reads each node's checker-proved type off `MTerm::ty()`. The build path is `Workspace -> elaborate_workspace -> MirWorkspace -> lower_workspace`; the public `lower_workspace*` entry points and their 18 call sites thread `&MirWorkspace`. The codegen-side type re-derivers `synth_with_extras` + `synth_arg_type` (and the `builtin_ail_type` / `builtin_effect_op_ret` mirror tables) are deleted — grep-clean. The three re-derivers the spec keeps until mir.2/mir.3 (`type_home_module`, `is_static_callee`, the second `infer_module_with_cross`) stay. The mechanical Term->MTerm match-arm conversion was straightforward and compiler-enforced. The substance was a set of producer-side correctness gaps that only surface once codegen reads `MTerm::ty()` and once the build path re-synthesises the post-mono AST through the canonical `synth` (which, unlike the old codegen, fully re-unifies). Each was root-caused against a failing e2e fixture: 1. `qualify_local_types` stripped fn-type modes (rebuilt `Type::Fn` with empty `param_modes` / `Implicit` `ret_mode`), so a monomorphised polymorphic intrinsic (`RawBuf.set`) lost its `Own` ret-mode and the owned temporary leaked at the call site. Made mode-preserving, like its sister `qualify_workspace_types` and `Subst::apply` ( |
||
|
Brummel
|
449df13c9c |
fix(check): preserve fn-type modes through Subst::apply
`Subst::apply` rebuilt every `Type::Fn` with `param_modes: vec![]` and `ret_mode: Implicit`, discarding the modes the surrounding type carried. That was harmless while nothing downstream of inference read fn-type modes — but it is a latent under-specification: substitution is meant to be type-preserving, and a fn type's modes are part of its identity (an `(own T) -> (own U)` contract is not the same contract as its borrow-returning sibling). Modes are positional over `params`, and substitution remaps each param type element-wise without changing the arity, so the mode lists stay aligned — preserving them is a clone, not a re-derivation. The typed-MIR boundary (milestone 0060) makes this load-bearing: a node type synthesised by `lower_to_mir::synth_pure` ends with `subst.apply`, and the codegen drop path reads `ret_mode == Own` off a call's callee node to decide RC trackability. With the modes stripped here, an Own-returning call read back `Implicit`, nothing was trackable, and a heap-allocated binder leaked at scope close. Preserving the modes restores that signal at the source rather than re-deriving it in codegen (which is exactly the re-derivation the milestone exists to delete). Safe by construction: `unify`'s Fn arm destructures modes with `..` and compares only params/ret/effects, so carrying modes through `apply` cannot change unification; and `Type`'s custom `PartialEq` already treats `Implicit` and `Own` as interchangeable (only `Borrow` is distinct), so equality outcomes are unaffected except where preserving a `Borrow` is the more correct answer anyway. Whole `ailang-check` suite stays green. RED pin: crates/ailang-check/tests/subst_preserves_modes.rs. |
||
|
Brummel
|
600565d356 |
fix(check): localize own-ADT type-args in free-fn mono; complete the lower_to_mir producer
Additive producer-side work for the typed-MIR milestone (spec
docs/specs/0060-typed-mir.md), surfaced while building the mir.1b
codegen-consumption switch. Nothing here consumes MIR — codegen is
untouched; the whole existing suite stays green and one new RED pin
goes green. The mir.1b codegen switch itself stays in the working tree.
Headline fix — own-ADT type-arg localization in free-fn mono:
monomorphise_workspace's free-fn arm normalises every observed
type-arg through Registry::normalize_type_for_lookup. That
normalisation is load-bearing for cross-module dedup and a stable
specialisation symbol name, but it qualifies a *defining-module-own*
ADT to `<module>.<T>`. synthesise_mono_fn_for_free_fn then substitutes
those qualified type-args into the polymorphic source signature and
body, which are in the module's own *bare* convention (the qualify
pre-pass deliberately leaves own-module type-cons bare — lib.rs:4358,
"the current module sees its own types bare"). The result is a
self-inconsistent specialisation: e.g. fold_left specialised with
accumulator b := List<Int> minted
((std_list.List<Int>, Int) -> std_list.List<Int>,
std_list.List<Int>, List<Int>) -> std_list.List<Int>
— a qualified accumulator parameter against a bare list argument and a
bare Lam body. check_workspace never hits this (it synths the pre-mono
bodies, bare throughout); the old codegen tolerated it via the
now-deleted synth_arg_type's loose type-tracking. The post-mono synth
re-entry in lower_to_mir does full unification and correctly rejects
it: `expected std_list.List<Int>, got List<Int>`.
Fix: localize_own_types (mono.rs) strips the `<defining_module>.`
prefix back to bare for the module's own ADTs, applied to the type-args
before they are substituted, so the synthesised signature and body stay
uniformly bare-own. The *stored* type_args remain normalised, so the
symbol name and the Phase-3 rewrite-cursor key on the same canonical
form — the byte-identity invariant between synthesis and rewrite is
untouched. The ClassMethod arm needs no analogue: it already
substitutes the un-normalised target.type_ (mono.rs:841).
Alternatives rejected: (a) drop the normalisation in the shared
apply_subst_and_normalize helper — breaks cross-module dedup, mints
duplicate specialisations under divergent names. (b) Make
lower_to_mir's synth treat `<owner>.T` and bare `T` as equal during
unification — papers over the producer inconsistency in the consumer
and contradicts the established own-types-stay-bare invariant the rest
of check relies on. (c) Qualify the whole synthesised def own->qualified
— fights lib.rs:4358 and would require the lower path to also flip,
cascading the convention change through synth.
Producer completeness (additive, mirrors existing guards):
- lower_module skips Type::Forall defs (lower_to_mir.rs), mirroring
emit_module's guard. Polymorphic defs are stale source kept for
round-trip; their bodies carry mono-rewritten call names for
specialisations that were never synthesised, so a synth re-entry
would hit UnknownIdentifier.
- lower_module lowers non-literal const bodies to typed MTerm, carried
in MirModule.consts (+ MirConst in ailang-mir). A non-literal const
(e.g. a List ctor expression) needs a typed body once the consumer
walk takes &MTerm; literal consts emit a global and need no body.
- MirModule gains `ast: Module`, populated by lower_module — forward
scaffolding the mir.1b codegen consumption reads for module
structure. Populated here, consumed by the in-tree switch.
Verification: full `cargo test --workspace` green (697 passed, 0 failed)
with the codegen-consumption switch stashed out, including the new pin
crates/ailang-check/tests/lower_to_mir_ty.rs::poly_free_fn_accumulating_into_own_adt_elaborates
(RED before the localize fix with the exact mismatch above, GREEN after)
and the full e2e corpus (the own-ADT-poly demos elaborate clean and run
through the old codegen). This proves the producer fix is
regression-free and the remaining 22 e2e reds in the working tree are
codegen-consumption defects, not producer defects.
|
||
|
Brummel
|
5b4eb766c3 |
glossary: bootstrap AILang nomenclature ledger
Add design/glossary.md (25 entries) and wire it in via the paths.glossary profile slot, making it standing reading for every skill and agent role. Built by the glossary skill's bootstrap procedure: five read-only glossary-extractor agents swept the prose surface (design/contracts, design/models, design/INDEX.md, docs/PROSE_ROUNDTRIP.md, and the 60 docs/specs in three blocks), reporting recurring domain-concept terms and their competing synonyms with frequencies. docs/plans was excluded as a derived execution artefact that mirrors spec/contract vocabulary rather than originating it. Three apparently-contested clusters were resolved by record-reality against the authoritative current docs (CLAUDE.md + design/): the retired .ailx extension yields to .ail; Form-A wins over "Form A" (13:2); round-trip wins over roundtrip (15:10). The one genuine nomenclature choice -- the canonical term for the structured hashable form -- was decided by the user: JSON-AST is canonical, .ail.json is its on-disk file. The local conformance check caught and removed a Form-B / JSON-AST Avoid-list collision before write. |
||
|
Brummel
|
1a1a68df8b |
plan: mir.1b — codegen consumes MIR (the atomic switch)
Second half of spec iteration mir.1, planned against the landed mir.1a
infrastructure (
|
||
|
Brummel
|
135f4ceed7 |
feat(check): mir.1a — typed-MIR types + post-mono lower_to_mir + elaborate_workspace (additive)
First half of spec iteration mir.1 (docs/specs/0060-typed-mir.md, plan docs/plans/0114-mir.1a-mir-infrastructure.md). Purely additive: a new leaf crate plus a post-mono lowering walk plus the front-end entry. Nothing consumes MIR yet — codegen and the CLI are untouched — so the whole existing suite stays green (102 test-result-ok lines, 0 failed; the one ignored is the #49 pin, which lifts at mir.4). mir.1b flips codegen to consume MIR and deletes the synth_with_extras mirror. What lands: - `ailang-mir` (new leaf crate, depends only on ailang-core): MTerm / MArg / MArm / MLoopBinder / MNewArg / Callee / Mode / StrRep / MirDef / MirModule / MirWorkspace, structurally complete over all 17 Term variants plus the dedicated Str split (MTerm::Str{lit,rep}). Every later-iteration annotation field exists now at a mir.1 default (App.callee = Indirect, MArg.mode/consume_count = Owned/1, Str.rep = Static, New.elem = None); the struct shape will not change across mir.2–mir.5, only the values. MTerm::ty() reads the carried type. - `ailang-check::lower_to_mir`: the post-mono walk. Carries no second type engine — it calls the canonical `synth` per node with fresh throwaway inference scaffolding (subst/counter/residuals/…), applies the substitution, and threads only the lexical locals/loop_stack, maintained by the same push/pop rules synth uses (Let, Loop, Lam, LetRec, and Match via synth's own type_check_pattern). This is the "one engine, not two" property the milestone buys, on the producer side. - `ailang-check::elaborate_workspace`: the single front-end entry — check (diagnostics gate) → desugar+lift → monomorphise → lower_to_mir — transcribed from the CLI build path. check_workspace stays for the diagnostics-only callers. - Three ty-fill pins (crates/ailang-check/tests/lower_to_mir_ty.rs) load the #51/#53/#49 witnesses as fixtures and elaborate the whole workspace, so lower_to_mir is exercised over the full prelude+kernel, not just the tiny witness. Two new witness fixtures (examples/new_{rawbuf_size_only,counter_user_adt}.ail); #49 reuses the existing loop_recur_str_binder_no_leak_pin.ail. Deviations from the plan, all forced by ground truth and verified against the live code before commit: - FnDef.export is Option<String>, not bool — dropped from MirDef (the plan's note allowed this fallback). Codegen reads FnDef.export directly, unaffected. - fn_param_types did not exist — lifted the param-type extraction into a shared pub(crate) fn and rewired both mono.rs sites (collect_mono_targets, collect_rewrite_targets) to it, so the param-type source cannot drift between mono and lower_to_mir. The helper unwraps a top-level Forall then reads Type::Fn.params — identical to the inner_ty extraction it replaces; the early-return non-fn guard is preserved at both sites. Behaviour-preserving (whole suite green; this is the only edit to existing mono behaviour). - lower_module mirrors two more mono.rs scaffolding rules the plan under-transcribed: skip intrinsic-bodied fns (they are signature-only; lowering them would hit synth's Term::Intrinsic guard) and seed env.current_module + env.globals + env.imports per module so synth's Var lookup resolves post-mono specialisations (e.g. compare__Int) in the prelude. - The #53 pin asserts the lowered (new Counter 42) init carries ty Counter, not that it is an MTerm::New: a monomorphic `new` over a user ADT is desugared to a dotted call `(app Counter.new 42)` before lower_to_mir runs (desugar.rs:1078-1098), so it lowers to MTerm::App. MTerm::New survives only for a polymorphic `new` carrying a written NewArg::Type (the #51 RawBuf path, covered by the rawbuf pin). The pin's intent — ty-fill correctness — is preserved; the structural variant is mir.2/mir.5 territory. - ailang-mir uses workspace-inherited Cargo fields for consistency with the existing crates (added it to [workspace.dependencies]). refs #49 |
||
|
Brummel
|
438a009b83 |
plan: mir.1a — typed-MIR infrastructure (additive half of spec mir.1)
First of two plans for spec iteration mir.1 (docs/specs/0060-typed-mir.md).
mir.1 is the project's largest iteration — a full codegen frontend
switch from walking &Term to walking &MTerm — and it does not fit one
bite-sized plan, because the switch is atomic: there is no compiling
intermediate between "all Term" and "all MTerm" (dual-path is
spec-forbidden). The MIR *producer* side, by contrast, compiles and
tests in isolation. So mir.1 is sequenced across two plans:
- mir.1a (this plan): the additive producer — new leaf crate
ailang-mir (MTerm/MArg/Callee/Mode/StrRep, structurally complete over
all 17 Term variants plus the Str split), ailang-check::lower_to_mir
(post-mono walk that calls canonical synth per node and maintains
locals/loop_stack exactly as synth does — no second type engine), and
ailang-check::elaborate_workspace (check → desugar+lift → mono →
lower_to_mir). Nothing consumes MIR yet; codegen untouched; whole
suite stays green; three ty-fill pins load the #51/#53/#49 witnesses
as fixtures and elaborate the full workspace (prelude included).
- mir.1b (next plan, against the landed types): flip codegen's walk to
&MTerm, delete synth_with_extras + synth_arg_type (9 call sites across
lib.rs/drop.rs/match_lower.rs), thread the 18 lower_workspace* callers,
switch the CLI build path to elaborate_workspace.
Both together satisfy the spec's mir.1 row. The split is plan
granularity, not a spec change — the spec's mir.1 deliverable is
unchanged.
Orchestrator design calls baked in (planner judgement, no user fork):
dedicated MTerm::Str{lit,rep} variant (Str-split installs the mir.4
hook at the #49 loop seed); emit_ir keeps &Module (builds MIR
internally — minimal blast radius); lower_to_mir scaffolding mirrors
mono.rs:771-816; Match-arm binding reuses synth's own type_check_pattern.
Recon (plan-recon) mapped the full deletion blast radius for mir.1b
(9 synth_arg_type sites, not the 3 the focus-hint assumed; 18
lower_workspace* callers). The two new witness fixtures
(new_rawbuf_size_only, new_counter_user_adt) were verified ail
check-clean during planning.
|
||
|
Brummel
|
8bc2972594 |
spec: typed-mir — place lower_to_mir post-mono, correct one-engine framing
plan-recon flagged that monomorphise_workspace runs between check and codegen, and codegen lowers the post-mono workspace. The first draft of this spec placed lower_to_mir as the final phase of check_workspace (pre-mono) and claimed "nothing re-walks the AST". That is wrong on both counts: - Built pre-mono, MIR would not carry the monomorphic specialisations monomorphise_workspace appends — exactly the post-mono bodies codegen emits today — reintroducing the check↔codegen gap this milestone exists to close. - synth is a pure `&Term → Type` function and the authoring AST carries no node-ids, so MIR cannot be a side-table read off a check pass; it is built inline by a walk. The honest property is therefore not "no second walk" but "no second *engine*": codegen's synth_with_extras is a hand-copied mirror of synth that drifts (every raw-buf bug is a drift point). The milestone replaces the mirror with a single post-mono call to canonical synth, materialised as MIR. Corrections: architecture diagram and data flow now run check → lift → monomorphise → lower_to_mir; lower_to_mir is wrapped by a new front-end entry elaborate_workspace that subsumes the lift+mono orchestration the CLI does today; check_workspace stays for the diagnostics-only path (`ail check`). Node count fixed 27 → 17 Term variants (ast.rs:436). Re-dispatched grounding-check after the edit (post-PASS edit invalidates the prior report) — PASS, all load-bearing post-mono assumptions ratified against live code (main.rs build order, mono.rs synth use, 17-variant Term, no node-ids). |
||
|
Brummel
|
207c63649f |
spec: typed-mir check→codegen boundary contract
Open a new milestone introducing a typed mid-level IR (`ailang-mir`) as the single artefact crossing the check → codegen boundary, so codegen re-derives nothing. Root cause (architect drift review): the boundary today carries no typed artefact. `crates/ail/src/main.rs:2293` calls `check_workspace`, discards the result, and hands `lower_workspace` the raw `Workspace`. `CheckedModule` (ailang-check/src/lib.rs:1283) carries only top-level (Type, hash). Codegen therefore re-runs four independent derivations of what check already proved — type synthesis (`synth_with_extras`), callee resolution (`type_home_module` ×3 + `is_static_callee`), uniqueness/mode (`infer_module_with_cross`, second run), and Str representation (`!is_str` recur gate). Every raw-buf bug (#43/#46/#47/#49/#51/#53, all "check-clean, build-divergent") is a point where one of those re-derivations disagreed with check and got patched one leg at a time. Honest framing: this is removal of a re-derivation architecture, not a RED→GREEN of crashing programs. #51 and #53 build today (patched by |
||
|
Brummel
|
02775b58ca |
test(codegen): mark the #49 Str-leg RED pin #[ignore] pending the representation spec
The Str-leg pin (committed |
||
|
Brummel
|
04f75c8b71 |
test(codegen): RED-pin #49 Str leg — heap-Str loop binder must not leak across recur
A heap-Str accumulator threaded through a `(loop …)`/`recur` as a loop binder leaks: each iteration's superseded `str_concat` slab is never RC-dec'd, even when the loop's final result is consumed. The fixture reports `allocs=4 frees=1 live=3` under AILANG_RC_STATS (stdout `xyyy` is correct). Root (data-flow traced): the `Term::Recur` arm in crates/ailang-codegen/src/lib.rs (~2131) gates its superseded-value dec behind `!is_str`. That exclusion exists because a `Str` loop seed may be a static-literal Str — a constexpr GEP into a packed-struct global with no rc_header (lib.rs:1612) — and `ailang_rc_dec` on a static pointer reads garbage at `payload-8` and aborts on underflow (rc.c:221, no runtime guard, per design/contracts/0011-str-abi.md). So the Str binder's prior heap slab is overwritten by a plain `store` with no dec. Distinct from the loop-RESULT scope-close drop ( |