iter embedding-abi-m2.1 (PARTIAL 4/9 + Boss spec-defect repair): ctx ABI + de-globalisation

Tasks 1-4 land fully review-green and are the cohesive shippable
subset (the per-thread embedding ABI, fully wired + proven):
- T1: FIXED-FIRST alloc-guard baseline pin (RED captured -> GREEN at T4).
- T2: runtime/rc.c gains ailang_ctx_t {alloc_count,free_count} +
  ailang_ctx_new/_free + __thread __ail_tls_ctx; the two increment
  sites become 'if (_ctx) _ctx->... else g_rc_...'. g_rc_* statics +
  ailang_rc_stats_atexit + the constructor RETAINED VERBATIM as the
  null-ctx (single-threaded executable) fallback. rc_accounting_tsan.c
  + driver: per-ctx 8x200000 tsan-clean (de-globalisation positive proof).
- T3: codegen Target::StaticLib forwarder gains a leading 'ptr %ctx'
  + one '@__ail_tls_ctx = external thread_local global ptr' decl +
  TLS save/store/restore around the BYTE-UNCHANGED internal call;
  _adapter/_clos + internal arg vector untouched (M1 decision held);
  doc-comment provisionality narrowed to the value/record layout.
- T4: build_staticlib rejects --alloc != rc (RC-only swarm artefact).

Boss adjudication of the orchestrator's Task-5 BLOCKED (spec-defect,
correctly surfaced not hacked): swarm.c's -DSHARED_CTX negative
control is structurally impossible — examples/embed_backtest_step.ail
is a non-allocating scalar kernel that never writes a ctx field, and
__ail_tls_ctx is __thread, so a shared-ctx scalar swarm has no
shared-memory write to race on (the spec's OWN item-1 honesty point).
The de-globalisation teeth belong to the item-1 rc_accounting harness
(shared ctx genuinely races on ctx->alloc_count; orchestrator
independently confirmed tsan exit 66). Spec amended in lockstep
(Goal coherent-stop, must-fail-axis item 2, Testing item 3,
Acceptance) so item 3's negative control is item-1's by the
de-globalisation-proof / capability-demo split; plan Task 5
restructured (swarm.c per-ctx capability demo only; negative-control
standing test relocated into the item-1 harness driver) + Task 3's
plan-transcription error (@ail_backtest_step ->
@ail_embed_backtest_step_step) corrected. This is a Boss
consistency-repair of an internally-contradictory clause whose intent
is already correctly realised in the same spec — not a redesign; no
brainstorm bounce. Task-5 working files discarded (corrected plan
reproduces them; a structurally-RED test must not enter main).
INDEX.md + final journal land at iter completion (re-dispatch [5,9]).

bench/orchestrator-stats/2026-05-18-iter-embedding-abi-m2.1.json

@@ -0,0 +1,14 @@
+{
+  "iter_id": "embedding-abi-m2.1",
+  "date": "2026-05-18",
+  "mode": "standard",
+  "outcome": "PARTIAL",
+  "tasks_total": 9,
+  "tasks_completed": 4,
+  "reloops_per_task": { "1": 0, "2": 0, "3": 0, "4": 0, "5": 0 },
+  "review_loops_spec": 0,
+  "review_loops_quality": 0,
+  "blocked_reason": "spec-ambiguous",
+  "blocked_task": 5,
+  "notes": "Tasks 1-4 fully green (zero review re-loops). Task 5 implemented faithfully but BLOCKED: shared-ctx swarm negative control is structurally unable to race for a non-allocating scalar kernel (@__ail_tls_ctx is __thread; no ctx field is written by the kernel). Real de-globalisation negative control already has teeth in Task 2's rc_accounting_tsan.c -DSHARED_CTX (tsan exit 66 at rc.c ailang_rc_alloc). Task 3 carried one DONE_WITH_CONCERNS plan-transcription-error fix (assertion symbol @ail_backtest_step -> @ail_embed_backtest_step_step to match the byte-unchanged @ail_<mname>_<fn> invariant). Tasks 6-9 not run (downstream resequencing is a Boss decision)."
+}