research: controlled shell call-shape diagnosis and wait verification (config-default wait campaign) - #69

Summary

Controlled shell call-shape diagnosis and wait verification — the full execution of plan docs/superpowers/plans/2026-10-16-rescue-shell-wait-shape-diagnostic.md under the unchanged 2026-10-05 spec (S1–S7, Q1–Q6, §8). Production scope: ZERO (prohibited-path diff vs 7fe8cbb is empty — skills/agents/scripts/hooks/marketplace/specs untouched). All changes are the research instrument (tools/shell-wait-probe/), its tests, and the dated evidence records (docs/qualification/rescue-shell-long-wait.md §15.1–§15.7).

Key results

• D0 (ledger 1/4): the A0 blocker did not recur — the fresh cell was admitted naturally by the existing grammar; managed-route long-wait observation measured end to end (route actual=named, 1 launch/1 send, original handle, empty poll 97.5 s, exit 0, sentinel matched).
• C0/C1 (ledger 3/4, D1 unspent cause-less): with --wait-policy config-default at 30000 ms the decisive poll yielded-pending at 31007 ms and the model continued the SAME cell via the sanctioned tools.wait continuation; at 240000 ms the work completed at 148.8 s without yielding. Identical instructions (normalized hashes), model (gpt-6.1-sol, discovered via authenticated catalog as the account's current default) and binary SHA-256 across arms.
• Adjudication v2: config-effect-observed on this single measured pair (adjudicationVersion: 2; v1 inconclusive/timing-borderline preserved as history — close-out correction 2 made the control band genuinely symmetric per user-commissioned review). Caveats binding: single pair, no repeat; no inner-cap/quantified/native-cancellation/production-readiness claim. Spec §8's production gates (quantitative long discriminator + repeat, reduced re-entry, exact terminal linkage, native pending-observation interruption with owned settlement) remain entirely outstanding.
• Shape diagnostic instrument: memory-only inspectCallShape classifier (canonical-exact architecture: pinned positive grammar + explicit variant set + one conservative unknown path; frozen schema v1 with the dated retention amendment), opt-in via --unclassified-shape-diagnostic 1 (requires the structural flag), bound into observer provenance digest v2. Fixture-proven (privacy/adversarial batteries); it had no real unknown cell to classify during D0 — recorded honestly.
• Research-only driver additions: --model <id> (bounded safe identifier, separate -m argv, requested recorded / effective stays null) and the task-4 pair above; legacy behavior byte-identical when omitted.

Verification

• Final source: npm run test:shell-research 594/594 (55 suites), selection 6/6, lint/typecheck/LF clean, git diff --check clean; full npm test at final source: exit 0 — 3397 tests, 3394 pass, 0 fail, 3 skipped (§15.7).
• Every record cites its tested revision; the run ledger, adjudication history (v1/v2) and honest limits are in report §15.5–§15.7.

Review history (all gates zero-findings before close)

• Task 1: self-review 6 rounds + codex review zero findings.
• Task 2 (classifier): 17 codex rounds — 37 findings on the original carried-context design → canonical-exact redesign (root-adjudicated) → 6 findings → round-17 APPROVE. Channel note: plain review died 4/4 on the post-redesign diff (reviewer-side node-repro wedge; our code fuzzed clean — 3000 mutations <500 ms); completed via adversarial-review with honest scope statements (recorded deviation, report §15.2).
• Corrections 1–2 + Task 4 (transport + adjudicator): each narrow-rechecked and gate-approved (task 4: 7 rounds 4→3→2→1→1→1→0).
• Branch-level close-out review: APPROVE, zero material findings.

Adjudicated deviations (all recorded in the report)

Canonical-exact redesign with disclosed coarsenings (§15.2); channel switch to adversarial-review (§15.2 channel record); round-2 timing-borderline semantics superseded by correction 2's symmetric band per user-commissioned review (§15.7); gpt-6.1-sol selected by authenticated read-only discovery per handoff primary path (§15.5).