docs: record failure-safe PI proof

.project-docs/30-worklog/tasks/20260824-pi-release-proof-3e725ac7.md

@@ -66,9 +66,10 @@
 
 ## Outcome
 
-- Final code candidate `0d26d17cfc5d32c91d6e25d919be1fabd94a605b`
-  closes all three P1 implementation/proof defects from the planner's second
-  review. Final-product proof now runs Main from packaged `app.asar`, starts the
+- Final code candidate `09841c8dbd7c58ce9ecda9dc49ea1c1b57759b67`
+  closes the three P1 implementation/proof defects from the planner's second
+  review and the two P1 defects from its third review. Final-product proof now
+  runs Main from packaged `app.asar`, starts the
   packaged `resources/pi-runtime/dist/cli.js`, receives a real `subagent` tool
   call from a controlled loopback provider, dispatches through the real
   `PiSubagentScheduler`, and starts a real ephemeral Pi child with exactly the
@@ -79,11 +80,22 @@
   `agent.settled`. Direct admission emits an explicit zero-duration queue wait;
   reopening the same session after a resource rebuild is correctly classified
   as warm.
+- `provider.first_event` now accepts only Pi's Provider-backed assistant
+  `message_start`, never `turn_start` or the local user-message lifecycle. The
+  controlled Provider delays its first response by 75 ms; the packaged smoke
+  asserts the source marker, strict timestamp ordering, and a measured delay of
+  at least 75% of that value, so the old zero-millisecond false milestone fails.
 - The 4+4 pressure proof now holds four real persistent Pi parents and four
   real ephemeral Pi children as eight distinct live OS processes while four
   write leases are active, proves the product UI remains interactive, and then
   proves all processes, provider requests, process-budget entries, child
   permits, dispatches, parents, and write leases return to zero.
+- Pressure cleanup now attempts every bounded cleanup step even when an earlier
+  step fails, keeps the global run handle until cleanup succeeds, and supports
+  idempotent retry. Unit tests cover one-shot failure and a hanging step; final
+  Windows and Linux products also inject a `parents.settle` failure into a real
+  4+4 run, observe the expected first failure, retry, and prove every counter
+  and PID returns to zero.
 - Real concurrent parent startup exposed a supported Windows failure in the
   shared managed provider catalog: simultaneous temporary-file renames could
   fail with `EPERM`. Same-path catalog writes are now serialized, with focused
@@ -109,50 +121,50 @@
 ## Verification
 
 - Exact pnpm `10.33.4` frozen install passed on Windows and WSL2 Ubuntu 24.04
-  x64; both final-product runs were bound to clean commit `0d26d17`. The
+  x64; both final-product runs were bound to clean commit `09841c8`. The
   lockfile change remains limited to the direct `@electron/asar@3.4.1`
   verifier dependency.
 - Candidate-final Windows checks passed: `pnpm run typecheck`; `pnpm run
   lint:check` with 0 errors and five pre-existing warnings; focused proof tests
-  with 6 files/25 passed/2 skipped; focused concurrency tests with 4 files/26
-  passed; `pnpm test` with 178 files/1506 passed/2 skipped; `pnpm run test:e2e`
-  with 24/24; and repeated `pnpm run build:vite` production builds.
-- Final Windows NSIS `release/Makelore-2.0.0-win-x64.exe` is 208,369,701 bytes
+  and cleanup tests passed; focused concurrency tests passed; `pnpm test` with
+  179 files/1508 passed/2 skipped; `pnpm run test:e2e` with 24/24; and repeated
+  `pnpm run build:vite` production builds.
+- Final Windows NSIS `release/Makelore-2.0.0-win-x64.exe` is 208,525,071 bytes
   with SHA-256
-  `C68DF05932055CF950A3491CB34CACB8A2D39F8F3D987B32FE4950621D1CC9B6`.
+  `396D6E8EC2BED60498DBC36CDA714176B629574914B85A5651C73DF8A63CECAA`.
   Its unpacked `app.asar` SHA-256 is
-  `F03278F6CF08B2D72F1518EFAA8338A92A8DDA0261AEEFCE82D93ACB9333F4D8`.
+  `872E8D1E7F22994EFF10353DF380016A198C59C145EBF63491E13DA1B1071B7F`.
   The post-installer-build final product proof passed from that ASAR with real
   Pi parent/child PIDs and complete cleanup. Windows structured reports are
-  `release/evidence/pi-smoke-win-real-workers.json`,
-  `release/evidence/pi-performance-win-real-workers.json`, and
-  `release/evidence/pi-product-proof-win-final-installer-build.json`.
+  `release/evidence/pi-smoke-win-final.json` and
+  `release/evidence/pi-performance-win-final.json`.
 - Windows formal five-sample performance passed every budget and recorded all
   nine cold/warm milestones with five samples each. Cold p95 was queue `0 ms`,
-  resources `32 ms`, spawn `14 ms`, RPC `718 ms`, session `4 ms`, accepted `1
-  ms`, agent start `1 ms`, first event `0 ms`, and settled `869 ms`; warm p95
-  was `0/11/8/624/1/1/0/0/64 ms`. Five 4+4 pressure samples had UI p95 `188
-  ms`; Renderer first-commit p95 was `33.466 ms`; exit p95 was `19 ms`.
-- WSL2 Linux x64 final product was rebuilt at exact clean commit `0d26d17`.
+  resources `30 ms`, spawn `18 ms`, RPC `750 ms`, session `4 ms`, accepted `2
+  ms`, agent start `1 ms`, Provider first event `151 ms`, and settled `1034
+  ms`; warm p95 was `0/10/9/660/1/2/0/144/146 ms`. Five 4+4 pressure samples
+  had UI p95 `140 ms`; Renderer first-commit p95 was `33.6862 ms`; exit p95 was
+  `16 ms`.
+- WSL2 Linux x64 final product was rebuilt at exact clean commit `09841c8`.
   AppImage/DEB/RPM SHA-256 values are respectively
-  `38c71302e26a0f7946ffcf72e23edd176b0abc1630da62ef25a808fddb6fdd5c`,
-  `7c4b29d819aba21c66abb1f41d62cd5b8aed24e44823e9eb875981133672f22b`,
-  and `cf6983f64a9d2c1f1080fabd9e1724914662e559843b9ceb4f11198a7312b917`.
+  `d0923c3d80c046c3e98750f7d55696e2753b710f9441d043a94768e0395a69c5`,
+  `930cc826993506f56544adc3c29fcddd993117c04e87332c81d031fd9d1ce070`,
+  and `81c1ceefbbe6314dec9c98b1441507458f23eea52c9acbf1360be0db20a88986`.
   Final ASAR SHA-256 is
-  `300dfeb32ded8c6fa764c6604f7d10c175dcbdc28e219169402f65a0dd50d7f7`.
-- Linux formal smoke passed Pi `0.84.2` production closure, 10,282-entry ASAR
+  `9ce9bce7a36e9d93a866b6982fb53c9e98f410d2dd92003fee2d3370deff7b15`.
+- Linux formal smoke passed Pi `0.84.2` production closure, 10,284-entry ASAR
   enumeration, zero product-owned OpenCode paths, actual final Pi processes,
   real packaged extension/subagent dispatch, 4+4 pressure, and zero cleanup.
   Reports copied beside the Windows evidence are
-  `release/evidence/pi-smoke-linux-wsl2-real-workers.json`,
-  `release/evidence/pi-performance-linux-wsl2-real-workers.json`, and
-  `release/evidence/pi-product-proof-linux-wsl2-real-workers.json`.
+  `release/evidence/pi-smoke-linux-wsl2-final.json` and
+  `release/evidence/pi-performance-linux-wsl2-final.json`.
 - Linux formal five-sample performance passed every budget and recorded all
   nine cold/warm milestones with five samples each. Cold p95 was queue `0 ms`,
-  resources `18 ms`, spawn `5 ms`, RPC `498 ms`, session `3 ms`, accepted `1
-  ms`, agent start `1 ms`, first event `0 ms`, and settled `602 ms`; warm p95
-  was `0/5/4/492/1/1/0/1/45 ms`. Five 4+4 pressure samples had UI p95 `142
-  ms`; Renderer first-commit p95 was `28.217 ms`; exit p95 was `9 ms`.
+  resources `14 ms`, spawn `6 ms`, RPC `480 ms`, session `2 ms`, accepted `2
+  ms`, agent start `0 ms`, Provider first event `114 ms`, and settled `740 ms`;
+  warm p95 was `0/6/5/467/1/1/0/111/112 ms`. Five 4+4 pressure samples had UI
+  p95 `115 ms`; Renderer first-commit p95 was `28.3212 ms`; exit p95 was `9
+  ms`.
 - Linux results are WSL2/WSLg final-product evidence, not an independent native
   non-WSL Linux desktop/compositor/distribution acceptance run. Controlled
   provider-shaped smoke issued six requests per protocol (distinct cold, warm,
@@ -161,11 +173,11 @@
 
 ## Follow-ups
 
-- Request planner re-review of `dc166a1..0d26d17`. The implementation evidence
-  now directly closes final packaged subagent dispatch, the single Main-owned
-  cold/warm timeline including zero-valued queue wait, and real
-  persistent-parent/ephemeral-child 4+4 pressure. The concurrent catalog-write
-  fix and its regression test are also included.
+- Request planner re-review of `ead9d1d..09841c8`. In addition to the earlier
+  packaged subagent, correlated timeline, and real 4+4 fixes, this range makes
+  Provider first-event timing response-backed with delayed-response assertions
+  and makes pressure cleanup bounded, failure-safe, and retryable with both
+  unit and final-product fault-injection evidence.
 - Do not start PI-160 or integrate the candidate while PI-150 remains blocked
   by the native non-WSL Linux platform evidence boundary. If the planner finds
   no new implementation defect, keep PI-150 as the active frontier until that