# iter cma.2 — Harness binary + four tasks + reference solutions + three integration tests

**Date:** 2026-05-12
**Started from:** 21c2d2cfaa999980f727a069792e3bd073c1f1e7
**Status:** DONE
**Tasks completed:** 13 of 13

## Summary

Stands up the sibling `harness/` standalone Cargo crate under
`experiments/2026-05-12-cross-model-authoring/`, mirroring the
out-of-workspace idiom from cma.1's `render/`. Six modules under
`harness/src/`: `strip_locations` (regex pass for form-asymmetric
location info), `pipeline` (subprocess wrapper around
`ail parse|check|build` plus the model's compiled binary, with a
fail-fast preflight on `ail` + `clang`), `ionos` (blocking reqwest
client with retry policy), `mock` (canned-response loader keyed by
cohort/task/turn), `scoring` (CSV + summary.md writers), `tasks`
(task-definition struct + loader). `main.rs` ties them into the
per-(cohort, task) loop. Four task definitions (`t1_add_three`,
`t2_length`, `t3_main_prints`, `t4_count_zeros`) with reference
solutions that all reach green through the actual ail+clang
pipeline. 13/13 tests pass: 5 inline unit tests under `--lib`
(strip_locations), 5 strip_locations integration tests against
verbatim recon-captured stderr fixtures, 1 verify_references
integration test driving every reference through the real pipeline,
1 mock_full_run end-to-end with eight rows and per-cohort artefact
trees, 1 budget_abort with the harness aborting cleanly under a
1500-token cap.

## Per-task notes

- cma.2.1: Bootstrap harness Cargo project + skeleton — created
  `harness/{Cargo.toml,.gitignore,src/lib.rs,src/main.rs}` plus
  six stub module files. Empty `[workspace]` table at the top of
  Cargo.toml keeps the crate out of the root workspace. Cargo.lock
  gitignored per spec. First `cargo build` succeeded with the
  expected unused-import warnings on the skeleton.
- cma.2.2: strip_locations module — written verbatim from the
  plan with 5 inline unit tests (json_pointer, byte_offset, line/
  column, Caused-by chain, passthrough). All 5 green on first run.
- cma.2.3: pipeline module — written verbatim from the plan plus
  a small implementer-phase repair: `ail check` enforces filename
  stem == module name, so `pipeline::run_pipeline` extracts the
  top-level `"name"` field from the JSON via a new
  `module_name_from_json` helper and renames the working file to
  `<module_name>.ail.json` before invoking `ail check`. Without
  this, every reference would have failed at the `ail check` step
  with "module name in file does not match expected name from
  path". See Concerns.
- cma.2.4: ionos module — written verbatim from the plan; added
  `thiserror = "1"` to `[dependencies]`. Compile-time warnings
  about three unused things on the `IonosClient` path (Auth, Usage,
  `last_err` assignment) are expected and flagged in the plan.
- cma.2.5: mock module — written verbatim; compile clean.
- cma.2.6: scoring module — written verbatim; compile clean.
- cma.2.7: tasks module — written verbatim; compile clean.
- cma.2.8: Author four task definitions + reference solutions —
  task.json files written verbatim from the plan text. Reference
  `.ail.json` files authored fresh against the cma.1 templates
  (hello.ail.json shape for t3; gc_stress.ail.json's polymorphic
  `List a` shape for t2 and t4; canonical chained `+` for t1's
  three-arg sum; canonical `eq` invocation pattern + branched `if`
  for t4's count_zeros). All four references compile clean and
  produce the expected stdout. Reference solutions intentionally
  omit `param_modes` because every parameter is Implicit and the
  canonical AILang form omits the field in that case — consistent
  with hello.ail.json, list.ail.json, eq_primitives_smoke.ail.json,
  and the explanatory doc in `master/examples/param_modes_all.ail.json`.
  See Concerns. Added `tempfile = "3"` to `[dev-dependencies]` for
  the `verify_references.rs` integration test. The test runs every
  reference through `pipeline::run_pipeline(Cohort::Json, ...)`
  end-to-end (ail parse-check-build-execute); passes when
  `AIL_BIN` is set to the release binary.
- cma.2.9: Wire main.rs end-to-end — written verbatim from the
  plan; added `tempfile = "3"` to `[dependencies]` (not only
  dev-deps) because `main.rs::run_one` uses it for per-turn
  workdir. Compile clean (modulo the three flagged warnings from
  ionos).
- cma.2.10: Five stderr fixtures + strip_locations integration
  test — all five fixture files written verbatim from the plan
  (which itself captured them from recon's run against current HEAD).
  Integration test asserts strip behaviour against each fixture;
  5/5 green on first run.
- cma.2.11: mock_full_run integration test + mock_full_run.json
  fixture — fixture authored by inlining each
  `master/tasks/<id>.reference.ail.json` as the `content` field
  for the green-path entries, plus `ail render` for the AILX-cohort
  inlines (verbatim AILX serialisations of the same references).
  Test runs the harness binary against the fixture; 8 rows total
  (4 tasks × 2 cohorts), `(json, t3_main_prints)` reaches green on
  turn 2 (matches the plan's intended cycle: broken-then-fixed),
  `(ailx, t1_add_three)` runs to the 5-turn limit (`(module garbage)`
  repeated). All artefacts (turn_N_program.ext,
  turn_N_request.json, turn_N_response.json,
  turn_N_check_stderr.txt, turn_N_run_stdout.txt) land in
  `per_cohort/<cohort>/<task>/`. 1/1 green.
- cma.2.12: budget_abort integration test — test verbatim from
  the plan, plus an implementer-phase repair to `main.rs`: with a
  tiny budget (1500), the outer loop broke out after consuming
  ~2400 tokens across tasks 1+2 (both green on turn 1), but
  neither row carried `final_status = budget_abort` because the
  inner `run_one` only marks BudgetAbort if the budget is exhausted
  *during* the turn loop. The plan's test assertion expects at
  least one row with `budget_abort`. Repair: after the outer loop
  breaks with `run_status="budget_exceeded"`, fill in the not-yet-run
  (cohort, task) pairs with synthetic `BudgetAbort` rows. mock_full_run
  remains green (the new code only runs on budget_exceeded).
  1/1 green. See Concerns.
- cma.2.13: README update + final sweep — appended "Running the
  harness" section to `experiments/.../README.md` verbatim from the
  plan. Full `cargo test` sweep: 5 (lib unit, strip_locations) + 5
  (integ, strip_locations) + 1 (verify_references) + 1
  (mock_full_run) + 1 (budget_abort) = 13/13 green. `git status`
  shows the README modification + the new harness/ tree + the new
  master/tasks/ files, all unstaged.

## Concerns

- **Pipeline filename-matching: minimal repair landed in `pipeline.rs`,
  not flagged in the plan.** `ail check` enforces filename stem ==
  module name; the plan's `run_pipeline` writes the model's program
  to `prog.{ext}`, which would have been rejected at every `ail
  check` invocation. The repair adds a 9-line helper
  `module_name_from_json` (reads JSON, returns top-level `name`) plus
  14 lines in `run_pipeline` that rename `prog.ail.json` ->
  `<module_name>.ail.json` between parse and check. Behaviour
  outside that rename is unchanged. The structural alternative
  (telling the model to name its module `prog`) would have been
  semantically wrong — the task description prompts say "module
  named t1_add_three" etc. Recommend a forward-fix sweep in cma.3
  if the structural decision (whether the pipeline or the prompt
  owns module naming) warrants spec attention.
- **`param_modes` omitted from every reference solution.** The plan's
  Step 8.2 text says "mode annotations on every parameter", but the
  templates the plan references (hello.ail.json, list.ail.json,
  eq_primitives_smoke.ail.json, etc.) all omit `param_modes`, and
  the `param_modes_all.ail.json` master example's own doc string
  documents "Implicit mode is the legacy default — `param_modes` is
  omitted from canonical JSON when every entry is Implicit". The
  references stay in canonical form. If the plan's intent was
  explicit `["implicit", "implicit", ...]` annotations on every
  reference, that is a follow-on amendment.
- **Budget-abort row-filling, minimal repair landed in `main.rs`,
  not flagged in the plan.** Without it, Task 12's test fails on
  the assertion `expected at least one budget_abort row`. Repair
  is gated on `run_status == "budget_exceeded"` so the green path
  is unchanged.
- **README "Total 8 passed" text mismatches Step 13.2's "13 passed"
  expectation.** The plan's README text (Step 13.1) advertises four
  test suites with "Total 8 passed" — counts the four integration
  suites but skips the 5 inline unit tests under `--lib`. Step 13.2
  expects 13 across `--lib` + 4 integration tests. The README text
  is reproduced verbatim per plan; the actual sweep produces 13.
  Recommend a docfix in cma.3 ("Total 13 passed across --lib + 4
  integration tests"), but the cma.2 plan locked the README copy.
- **`AIL_BIN` env required for `verify_references.rs` and
  `mock_full_run.rs` to run from a fresh shell.** These tests shell
  out to `ail`; without `ail` on PATH the preflight bails. The
  README documents this. CI integration is out of scope.

## Known debt

- The harness has not yet been live-fired against IONOS (deferred
  to cma.3 per spec). The retry/backoff path in
  `ionos::IonosClient::post` has no test coverage in cma.2 (mock
  mode bypasses it entirely).
- Three compile warnings in `ionos.rs` (`Auth` and one `Usage` are
  read in tests/other code; `last_err` second assignment is
  intentional dead-store for retry-loop symmetry). Not addressed.

## Files touched

Modified (1):
- `experiments/2026-05-12-cross-model-authoring/README.md`

New — harness crate (8 source, 4 integration tests, 6 fixtures, 1
manifest, 1 .gitignore):
- `experiments/2026-05-12-cross-model-authoring/harness/Cargo.toml`
- `experiments/2026-05-12-cross-model-authoring/harness/.gitignore`
- `experiments/2026-05-12-cross-model-authoring/harness/src/lib.rs`
- `experiments/2026-05-12-cross-model-authoring/harness/src/main.rs`
- `experiments/2026-05-12-cross-model-authoring/harness/src/strip_locations.rs`
- `experiments/2026-05-12-cross-model-authoring/harness/src/pipeline.rs`
- `experiments/2026-05-12-cross-model-authoring/harness/src/ionos.rs`
- `experiments/2026-05-12-cross-model-authoring/harness/src/mock.rs`
- `experiments/2026-05-12-cross-model-authoring/harness/src/scoring.rs`
- `experiments/2026-05-12-cross-model-authoring/harness/src/tasks.rs`
- `experiments/2026-05-12-cross-model-authoring/harness/tests/strip_locations.rs`
- `experiments/2026-05-12-cross-model-authoring/harness/tests/verify_references.rs`
- `experiments/2026-05-12-cross-model-authoring/harness/tests/mock_full_run.rs`
- `experiments/2026-05-12-cross-model-authoring/harness/tests/budget_abort.rs`
- `experiments/2026-05-12-cross-model-authoring/harness/tests/fixtures/check_unbound_var.stderr`
- `experiments/2026-05-12-cross-model-authoring/harness/tests/fixtures/check_type_mismatch.stderr`
- `experiments/2026-05-12-cross-model-authoring/harness/tests/fixtures/check_bare_xmod.stderr`
- `experiments/2026-05-12-cross-model-authoring/harness/tests/fixtures/check_schema_missing_field.stderr`
- `experiments/2026-05-12-cross-model-authoring/harness/tests/fixtures/parse_unclosed.stderr`
- `experiments/2026-05-12-cross-model-authoring/harness/tests/fixtures/mock_full_run.json`

New — master tasks (4 task.json + 4 reference.ail.json):
- `experiments/2026-05-12-cross-model-authoring/master/tasks/t1_add_three.task.json`
- `experiments/2026-05-12-cross-model-authoring/master/tasks/t1_add_three.reference.ail.json`
- `experiments/2026-05-12-cross-model-authoring/master/tasks/t2_length.task.json`
- `experiments/2026-05-12-cross-model-authoring/master/tasks/t2_length.reference.ail.json`
- `experiments/2026-05-12-cross-model-authoring/master/tasks/t3_main_prints.task.json`
- `experiments/2026-05-12-cross-model-authoring/master/tasks/t3_main_prints.reference.ail.json`
- `experiments/2026-05-12-cross-model-authoring/master/tasks/t4_count_zeros.task.json`
- `experiments/2026-05-12-cross-model-authoring/master/tasks/t4_count_zeros.reference.ail.json`

## Stats

bench/orchestrator-stats/2026-05-12-iter-cma.2.json