feat(context-providers): preregister A/A noise arms and recoverable failure diagnostics

[sarthakagrawal927]

Closes the two Stage 1 gaps recorded in benchmarks/context-providers/README.md so a later Stage 2 can be planned. Harness only — this PR executes no benchmark run and publishes no result. Stage 2 (and any paid, hosted, or public run) still needs its own explicit approval per #159.

Preregistered A/A noise arms

• The planner schedules independent treatment-vs-treatment A/A arms beside the A/B crossover (--aa-repetitions; default 2 at the full stage, opt-in at feasibility). A full-stage plan declaring none is blocked with aa-schedule-missing; 1 repetition is rejected outright.
• Every A/A arm gets its own workspace, agent session, and tool-configuration identity, so A/A noise cannot be derived by relabelling A/B repetitions — the projector rejects any attempt whose comparison/arm disagrees with the schedule, and family_qualified now requires complete A/A pairs.
• Plans without A/A arms keep their previous identity and counts, so the pinned Stage 0 and Stage 1 plan identities are unchanged.
• Cost derivation and attempt bounds account for the doubled treatment attempts, so a Stage 2 plan re-states its own cost for re-approval.

Recoverable failure evidence

Stage 1 could not explain its two treatment agent failures because the immutable receipt keeps only bounded hashes. The runner now also writes redacted, bounded stdout/stderr to <run-root>/diagnostics/attempt-<n>.json, verified against the receipt's stdout_sha256/stderr_sha256 before retention. Receipts are unchanged; a hash mismatch fails the run rather than storing unbound output.

Verification

Run in a clean worktree at origin/main with the patch applied, compared against the same worktree without it:

Check	origin/main baseline	This branch
pnpm test:corpus-contracts	61/61 pass	67/67 pass
pnpm test:context-provider-plan	—	11/11 pass
pnpm exec biome check .	clean (776 files)	clean (776 files)
pnpm quality:cycles	clean	clean
pnpm quality:duplication	exit 0	exit 0
pnpm knip:strict	fails (pre-existing)	byte-identical output to baseline

Refs #159.

[sarthakagrawal927]

Not merging — quality:complexity fails, and the change causes it

CI lint-and-typecheck → Code health step, verbatim (run 32602737946):

> code-reviewer@1.1.2 quality:change-size
Change-size gate: PASS (11/40 files, 809/4000 additions, 889/6000 gross lines)

> code-reviewer@1.1.2 knip:strict
Configuration hints (82)      <- unchanged from the origin/main baseline

> code-reviewer@1.1.2 quality:complexity
scripts/agent-task-corpus/context-provider-evaluation.mjs:182 projectContextProviderEvaluationBundle has complexity 38 (new violation)
scripts/agent-task-corpus/context-provider-plan-cli.mjs:40 parseArgs has complexity 28 (baseline 25)
scripts/agent-task-corpus/context-provider-run-cli.mjs:27 runContextProviderPlan has complexity 21 (new violation)
scripts/agent-task-corpus/contracts.mjs:1879 validateContextSchedule has complexity 34 (new violation)
 ELIFECYCLE  Command failed with exit code 1.

scripts/check-changed-complexity.mjs grandfathers a changed function at its origin/main score but caps everything at max(20, previous) (biome.code-health.json → noExcessiveCognitiveComplexity.maxAllowedComplexity: 20). Three of these four functions were under 20 on origin/main and this branch pushes them over; parseArgs regresses 25 → 28. There is no accepted-debt allowlist to record an exception in, so the gate can only be satisfied by extracting the new branching.

Everything else is green — see the verification table in the PR description. Leaving this open rather than merging or force-merging: reducing cognitive complexity inside the A/A scheduling, projection, and schedule-validation paths is a real refactor of correctness-bearing statistics code and should be a deliberate follow-up, not a rubber stamp.

Suggested extractions, smallest first:

• contracts.mjs validateContextSchedule — split the per-entry A/A vs A/B key/group derivation into a helper alongside the existing validateContextScheduleArm.
• context-provider-evaluation.mjs projectContextProviderEvaluationBundle — the A/B pair loop and the A/A pair loop are independent; each can become its own function returning runs.
• context-provider-run-cli.mjs runContextProviderPlan — the new diagnostics write is a natural extraction point.
• context-provider-plan-cli.mjs parseArgs — the value-taking flag list is now long enough to be a table lookup.

[sarthakagrawal927]

Landed on main as e7ccfc3 rather than merged here.

Why the merge button could not be used: I force-pushed main earlier to purge private-repository artifacts from history, which rewrote every commit SHA. This branch still points into the orphaned history, so GitHub computed the diff across two divergent ancestries — 711 commits each way, 100 files, and none of the files this PR's own description names. The change itself was always the 11 files it claimed. That breakage is mine, not this branch's.

Review notes, all acted on before landing:

Four functions were above the repo's cognitive complexity ceiling of 20 — projectContextProviderEvaluationBundle at 38, validateContextSchedule at 34, parseArgs at 28 (from a baseline of 25), runContextProviderPlan at 21. The verification table in the description covers lint, cycles, duplication and both test suites but not quality:complexity, which is why none surfaced. All four are split in e7ccfc3: schedule validation into per-entry/key-derivation/tally steps, the projector into separate A/B and A/A collectors, the flag chain into a table, and the runner preflight into plan-approval and adapter-pinning halves.

The A/A design holds up on reading. Separate workspace, session and tool-configuration identity per arm is the right call — deriving A/A noise by relabelling A/B repetitions is precisely what the measurement exists to rule out, and the projector rejecting a mismatched comparison/arm closes that off structurally rather than by convention.

67/67 corpus contracts and 11/11 plan tests pass, matching what you reported. Complexity, duplication, cycles, docs, change-size and lint are green on main.