Record continuation accounting baseline and retained live failures

doc/plans/2026-09-22-continuation-accounting-baseline.md

@@ -29,7 +29,7 @@ new paid productive cases use public questions and responses instead.
 | Repeated missing disposition | Existing repair episode unit/DB tests; complete sweep after exhaustion | Initial turn plus exactly two legacy repairs; visible board recovery, blocked task |
 | Comments/confident prose/tool calls without state change | Quiet/noisy exhausted repair and failure variants; native event replay | Quiet versus three distinct attributed misleading comments per run; no extra allowance |
 | Late approval/pause/spending/ownership gate | Second repair delayed, new gate, service restart, same debit; existing admission tests cover both runtimes | Approval created by first repair owns wait; acceptance causes one completion; Stop before second repair dispatch |
-| Restart while a repair is scheduled | Durable episode/dispatch tests | Controller restart before delayed repair; identical run IDs and episode counters |
+| Restart while a repair is scheduled | Actual delayed-retry promotion as well as the repair gate; retain episode and debit | Controller restart before delayed repair; identical run IDs and episode counters; fail with the persisted cancellation reason if promotion cancels it |
 | Duplicate/out-of-order handling | Concurrent retry/recovery tests and native consumed-repair replay with commentary/tool events | Exact run counts, document revisions and retained IDs; exhaustive races stay deterministic |
 | No reset after exhaustion; legitimate new request | Existing heartbeat exhausted failure versus actual new user request | No user message during exhaustion; a real approval resumes its owned path |
 | Missing or misleading evidence | Calibrated oracle rejects missing state, wrong documents, overwritten steps, fabricated progress, extra runs, lost receipts, decline, executed Stop | Every live result uses that oracle and existing screenshot/evidence/billing pipeline |
@@ -64,3 +64,11 @@ spending gates and timing races use deterministic fixtures rather than artificia
 spending money or hoping for a real outage. The original native compatibility
 probes, blank-route reload finding and duplicate-response finding remain in the
 [retained backlog](2026-09-22-legacy-continuation-authority.md).
+
+## Recorded baseline
+
+The [measurement report](../../tests/lifecycle-baseline/CONTINUATION-ACCOUNTING-2026-09-22.md)
+retains the deterministic inventory and each paid campaign without regrading.
+The new suite exposes shared allowance accounting in both directions and a
+delayed-repair promotion mismatch. These are production follow-ups; the test
+slice leaves their intended-behavior assertions enabled and failing.

doc/plans/2026-09-22-legacy-continuation-authority.md

@@ -123,8 +123,8 @@ Keep this section even if this slice incidentally resolves some assertions.
 | Commentary counts as progress | 1 | Retain for progress-evidence audit; no comment-count budget resets in this slice |
 | Wording selects liveness versus handoff path in heartbeat | 1 | Retain separately in results; must be covered while closing the current shared authority boundary |
 | Native autonomous continuation compatibility probes | 2 | Replace with reachable public-contract tests; `same_agent` injection is not a current model-facing defect |
-| Productive continuation versus repair/failure budgets | Coverage boundary | Separate audit after this slice; do not infer progress from text/event counts |
-| Blank task route after reload | New live observation | Investigate browser bootstrap/rendering; seen after successful persisted lifecycle in both runtimes; campaigns 35744074887 and 35747200170 |
+| Productive continuation versus repair/failure budgets | Three new findings | [Accounting baseline](2026-09-22-continuation-accounting-baseline.md) exposes delayed repair spending infrastructure allowance, infrastructure retries spending productive allowance, and delayed repair cancellation during promotion. Test-only slice; production fixes remain open |
+| Blank task route after reload | Repeated live observation | Investigate browser bootstrap/rendering; campaigns 35744074887 and 35747200170, plus accounting campaign 35813099816 after two correct legacy steps with a pending question. Empty screenshot retained; full quiet productive case remains unverified |
 | Duplicate legacy response | One later live observation | Campaign 35805477715 legacy work-mode neutral: provider issued two successful PATCH calls, second escaping an underscore. Retain as model behavior evidence; exact-once matcher rejected it |
 | Missing-comment policy ownership | Coverage boundary | Existing typed retry policy remains; review separately if consolidating all post-run compliance into one disposition contract |
 | Maintain CI and prepare App/Evals review | Pending | Promote fixed invariants into maintained cheap gates; paid suites remain explicit |

tests/lifecycle-baseline/CONTINUATION-ACCOUNTING-2026-09-22.md

@@ -0,0 +1,138 @@
+# Continuation accounting baseline — September 22, 2026
+
+This is a test-first slice. Production scheduling/authority code is unchanged.
+The [scenario matrix](../../doc/plans/2026-09-22-continuation-accounting-baseline.md)
+and executable inventory cover separate allowances, false progress, late gates,
+restart and duplicate handling. There are 67 additional deterministic tests
+and eight new explicit-only real-provider Product E2E cells.
+
+## Deterministic measurement
+
+Source: `379a70315e47de1ac2c8e4eb452eb3121401c5c7`, no tracked source changes at measurement.
+Raw evidence: `.lifecycle-baseline/2026-09-23T03-04-36.107Z/`.
+[Saved inventory and results](baselines/2026-09-22-continuation-accounting.json).
+
+| Layer | Passed / total | Failed |
+|---|---:|---:|
+| Unit | 413 / 414 | 1 |
+| Runner | 184 / 184 | 0 |
+| Database/heartbeat integration | 371 / 376 | 5 |
+| Grading | 85 / 85 | 0 |
+| Total | 1,053 / 1,059 | 6 |
+
+No skipped tests or missing inventory evidence. Four newly failing tests
+identify three production problems:
+
+1. **Delayed repair spends infrastructure allowance.** Immediate first repair
+   has no scheduled retry debit and passes. Delayed repair 2 has
+   `scheduledRetryAttempt: 2`, `scheduledRetryReason: issue_disposition_repair`.
+   `executionFailureRetryCount` interprets it as two prior failures, so a first
+   safe infrastructure failure is already exhausted. One pure assertion and one
+   actual scheduler assertion expose this.
+2. **Infrastructure retries spend productive allowance.** A run resumed after
+   two infrastructure retries reaches its first typed max-turn continuation.
+   The scheduler reads the same counter as two consumed productive attempts and
+   returns exhausted. The test expects productive attempt 1.
+3. **The delayed second repair is cancelled during promotion.** The legacy
+   repair dispatch gate accepts the persisted episode, but the shared dispatch
+   path compares its fingerprint against a different disposition fingerprint.
+   It cancels the unstarted repair with
+   `issue_disposition_repair_superseded`. The new database test reproduces this
+   through actual due-retry promotion after recreating the controller service.
+   The first live exhaustion run also retained this exact cancellation, with
+   only two provider executions and no final exhaustion checkpoint.
+
+The other two failures are the previously retained native `same_agent`
+compatibility probes. They inject an unsupported current public continuation
+shape and do not establish a current model-facing defect. The new paid native
+productive tests use reachable question/response contracts.
+
+The tests also prove that twenty comments or a large raw tool count do not
+replenish exhausted legacy repairs, including through the full recovery sweep
+while the issue is still active. A new approval, company budget hard-stop,
+agent pause or reassignment during the second repair delay blocks dispatch
+without another debit. Native replay with twenty commentary/tool events does
+not resubmit consumed disposition repair. Existing gates and duplicate/restart
+regressions remain linked in the inventory.
+
+## Other verification
+
+- Product E2E support: 474/474 passing, including 31 new accounting support tests.
+- Inventory/report support: 4/4 passing.
+- E2E and lifecycle baseline typechecks pass.
+- Server and runner package typechecks pass; server preparation also completes
+  the runner build and generated-contract checks.
+- Full general repository tests, recursive typecheck and root build were not
+  rerun: this slice changes only tests, harness and documentation. This is not
+  a claim of repository-wide green or PR readiness. The new intended-behavior
+  failures remain enabled and will fail their normal test lane until fixed.
+
+The [initial deterministic measurement](baselines/2026-09-22-continuation-accounting-initial.json)
+at `ed53efb1e0e363b382b5101340259ad3cab34307` is retained separately: 1,051/1,056.
+The later total adds the promotion regression and two harness calibration tests.
+
+## Live measurement
+
+All cells use qualified Codex `gpt-5.6-sol`, real Chromium, and isolated local
+server/database fixtures in parallel GitHub Actions jobs. Original machine
+grades, source/definition fingerprints, cleanup and billing are retained for
+each campaign; unavailable cost is not zero spending.
+
+First campaign: [35811412940](https://github.com/paperclipai/paperclip/actions/runs/35811412940),
+source `9cd7a585f5c69dc327b7d90534ad51be89ae5bb7`. All eight cells failed.
+This measurement is not regraded. Initial fixtures requested three identical
+comments, which the product deduplicates per run; rich-text serialization also
+escaped underscores in literal test tokens. Screenshot names did not meet the
+existing evidence allowlist, so the public report publisher rejected the bundle.
+Legacy quiet productive behavior passed every causal check before packaging
+failed. Approval and Stop lifecycle checks also passed while their comment
+perturbation checks failed. Both exhaustion cells failed to reach the expected
+exhaustion checkpoint; retained API state identifies the cancelled delayed repair.
+
+Corrected fixtures use three distinct numbered comments, plain alphanumeric
+tokens and existing allowed screenshot names. A packaging test verifies those
+screenshots survive. Cancellation fails promptly with its typed reason and the
+last observed state, instead of waiting for the overall deadline. These are
+test-harness changes, not changes to production behavior.
+
+Corrected campaign: [35813099816](https://github.com/paperclipai/paperclip/actions/runs/35813099816),
+source `379a70315e47de1ac2c8e4eb452eb3121401c5c7`:
+**5/8 passed, 3 failed**, no missing cells or retries, all eight evidence packages
+valid and cleanup successful. The corrected public report published successfully.
+
+| Cell | Result | Observed behavior |
+|---|---|---|
+| Legacy productive, quiet | Fail | After two successful runs, two exact revision-one records, and a pending next question, a task-route reload renders a blank page. `task-chat-thread` never appears within 30 seconds. Remaining three steps are unverified |
+| Legacy productive, noisy | Pass | Five productive runs, five exact revision-one documents, four real answered questions, no repair/failure debits |
+| Native productive, quiet | Pass | Same five-step causal assertions |
+| Native productive, noisy | Pass | Same five-step causal assertions despite misleading attributed comments |
+| Legacy exhaustion, quiet | Fail | Second repair cancelled before dispatch with `issue_disposition_repair_superseded` |
+| Legacy exhaustion, noisy | Fail | Identical typed cancellation despite the additional comments |
+| Legacy repair then Stop | Pass | Delayed second repair cancelled by the operator before dispatch; remains stopped after controller restart and scheduled due time |
+| Legacy repair then approval | Pass | Pending approval owns the wait; real browser acceptance produces exactly one successful completion without repair debt |
+
+The blank-page screenshot is empty. This matches the symptom of the
+retained task-route reload finding; it does not establish the underlying UI
+cause, and it is not counted as a completed productive case. The successful
+native noisy final screenshot shows Done and all five document cards at revision
+one. Both exhaustion checkpoints retain the scheduled and restarted run IDs;
+their last observations include the cancelled unstarted repair and a board-owned
+`active_run_watchdog` action instead of completed disposition-repair exhaustion.
+
+- [Published corrected report](https://d1p6rlowie26tp.cloudfront.net/runner-e2e/campaigns/gha-35813099816-1/index.html)
+- [Both immutable campaign measurements, fingerprints, grades and billing](baselines/2026-09-22-continuation-accounting-live.json)
+
+Billing is partial in both campaigns. Provider-reported zero totals do **not**
+establish free execution: the first campaign records token usage for 29/32 run
+records and reported cost for 10; the corrected campaign records usage for 26/29
+and reported cost for 10. Each includes three cancelled, unstarted records.
+The saved billing fields preserve the original coverage and cost status.
+
+## Next production slice
+
+Fix allowance separation and the delayed-repair promotion contract against the
+enabled deterministic failures, then rerun the same eight-cell suite. Keep the
+blank-route finding, unsupported native compatibility probes and earlier backlog
+distinct. This baseline establishes assertions and observed behavior; it does
+not claim the product is already robust or that every supported provider is
+qualified.

tests/lifecycle-baseline/README.md

@@ -150,3 +150,5 @@ E2E scores, or claim full qualification from a partial selection.
 September 22 follow-up: [legacy continuation implementation and verification](LEGACY-CONTINUATION-2026-09-22.md), including preserved failed campaigns and the remaining backlog.
 
 The [continuation accounting matrix](../../doc/plans/2026-09-22-continuation-accounting-baseline.md) adds ACCT-01 through ACCT-04 for separate allowances, false progress, late gates and restart/replay. Its real-provider companion is the explicit-only `continuation-accounting` Product E2E suite.
+The [September 22 measurement](CONTINUATION-ACCOUNTING-2026-09-22.md) records the
+enabled failures and preserves both initial and corrected live campaigns.

tests/runner-e2e/CONTINUATION-ACCOUNTING.md

@@ -36,3 +36,7 @@ See the [scenario matrix](../../doc/plans/2026-09-22-continuation-accounting-bas
 for deterministic fault, replay, spending and ownership coverage. Infrastructure
 failures are not induced in paid cells; their cross-lane allowance semantics are
 covered by actual scheduler tests with controlled failures.
+
+Results and follow-ups are retained in the [accounting measurement report](../lifecycle-baseline/CONTINUATION-ACCOUNTING-2026-09-22.md),
+including the original failed fixture campaign. A new measurement never replaces
+or regrades an earlier campaign.