Integration CI: tighten the -O2 inventory and fix timer.c for GCC 14+ (E4) #75

Activity

1. fd-shockwave-bot

   added this to the M5 — Establish host interchangeability milestone on Oct 20, 2026
2. fd-shockwave-bot

   added a parent issue on Oct 20, 2026

   • [Epic][M5] Prove host interchangeability while preserving the reference baseline #35
3. fd-shockwave-bot commented on Oct 20, 2026

   

   fd-shockwave-bot

   on Oct 20, 2026

   Last edited by fd-shockwave-bot

   Author

   More actions

   SUPERSEDED by the handover at #75 (comment) (PR #76 rescoped, candidate 303103d2). Kept for history.

   Handover (2026-10-20, fd-shockwave-bot)

   • Branch / commit: ci/integration-green at 2c4cc5ae. Base: claude/eech-campaign-native-spike-tjkj6g at 254ded56. The brief named b4d43ea6, but 254ded56 (GCC 12 pin, wine) landed before this work started, so this Task works forward from it. PR eech-campaign: tighten the -O2 global inventory; E4 timer.c callback fix for GCC 14+ (#75) #76 → integration branch; Task Integration CI: tighten the -O2 inventory and fix timer.c for GCC 14+ (E4) #75 under Epic [Epic][M5] Prove host interchangeability while preserving the reference baseline #35 (M5).
   • Scope / gate: CI and build hygiene only. No campaign, engine C, golden output, baseline, tolerance or pack.json change. This puts M5's refactoring under checks that actually run: the workflow had never passed (0 of 28 runs).
   • Root causes:
     • b4d43ea6, x86_64: timer.c:274. The original passes a DWORD-parameter callback where the compat shim declares LPTIMECALLBACK with DWORD_PTR parameters. GCC 14+ makes that an error. 254ded56 fixed it with GCC 12, not this PR.
     • b4d43ea6, windows-cross: "No such file": the cargo runner wine was not installed (only wine64). Also fixed by 254ded56.
     • 254ded56, x86_64 Lints: clippy 1.99's new chunks_exact_to_as_chunks in eech-map. This PR uses as_chunks::<2>().
     • 254ded56, x86_64 EECH_C_OPT=2: GCC 12 -O2 elides two write-only statics, old_num_map_z_sectors and update_data, and the inventory's "gone" check failed on them. This PR marks exactly those two manifest entries [elidable]. An optimised kernel may omit only them. At -O0 all symbols must exist, and the unclassified check is unchanged.
   • Red / green:
     • Red: reproduced in CI run 39428612290 (not locally: no Linux toolchain, and the Docker daemon is down).
     • Green: CI run 39430643777, 5 of 5 jobs successful (x86_64 incl. Lints, x86_64 OPT=2, i686, i686 original-stack, windows-cross). every_writable_global_is_classified ... ok in each job.
   • Regression: python tools\regression-pack.py --retained-only exits 0, with all 14 retained checks PASS (verdict "INCOMPLETE (retained tier only)", by design). A broader Windows runtime regression was not run and is not warranted: nothing in eech_dc.dll or eech-world.exe changed, and no crate depends on eech-map.
   • Boundaries: none discovered or changed. Baseline differences: none.
   • Limitations:
     • The [elidable] evidence covers GCC 12 -O2 only.
     • The windows-cross inventory may skip silently if nm cannot spawn under Wine (pre-existing).
     • regression-pack.py crashes on relpath when --out is on another drive.
   • Lead decisions requested:
     1. Whether to add a timer.c engine patch (callback parameters to DWORD_PTR; behaviour-neutral on i686; the shim never calls back) so the build no longer depends on the GCC 12 pin.
     2. Whether to enforce or update the declared workspace rust-version = "1.80". It is not inherited, and eech-map needs 1.88 now (1.87 before this PR).
   • Review status: formal review requested from @fd-starscream-bot. Not approved; not merged.
   • Next bounded action: address the lead's review feedback and decisions.
4. fd-shockwave-bot

   changed the title [-]Integration CI green at b4d43ea6[/-] [+]Integration CI: tighten the -O2 inventory and fix timer.c for GCC 14+ (E4)[/+] on Oct 20, 2026
5. fd-shockwave-bot commented on Oct 20, 2026

   

   fd-shockwave-bot

   on Oct 20, 2026

   Last edited by fd-shockwave-bot

   Author

   More actions

   SUPERSEDED by the handover at #75 (comment) (candidate aa110ffa). Kept for history.

   Handover (2026-10-20, fd-shockwave-bot), superseding the earlier handover

   • Branch / commit: ci/integration-green at 303103d2. It is merged with the integration head d2d4792a (94432391), and the integration branch has not moved since. Commits: 2c4cc5ae, 94432391 (merge, inventory tightening), 303103d2 (E4).

   • Scope (rescoped, user-approved):

     1. Inventory tightening: at EECH_C_OPT ≠ 0 only the two [elidable] globals may be absent.
     2. E4: timer.c's timeSetEvent callback and user data take DWORD_PTR.

     CI and build hygiene only; no campaign behaviour change.

   • Characterised: the generated code for timer.c with and without E4, compiled with GCC 12 and the engine's flags. i686 is identical. On x86-64 the only difference is that the pointer is no longer truncated (the headless timeSetEvent never calls back).

   • Red / green for timer.c:

     • Red: GCC 15 in CI run 39426483644 (b4d43ea6). Locally, on 94432391 without E4: GCC 14.2.0 (rust:trixie) and GCC 15.3.0 (gcc:15), both timer.c:274:66: error: … [-Wincompatible-pointer-types].
     • Green: the same two compilers with E4 build OK. The full x86_64 cargo test passes on GCC 14.2.0. The -O2 inventory passes 2 of 2 on GCC 14.2.0 and 15.3.0.

   • CI: run 39440605895 at 303103d2, 5 of 5 green (x86_64 incl. Lints, x86_64 OPT=2, i686, i686 original-stack, windows-cross).

   • Regression: python tools\regression-pack.py --retained-only exits 0, with 15 of 15 retained checks PASS. (The earlier handover said 14; it is 15.)

   • Boundaries: none discovered or changed. Baseline differences: none.

   • Limitations:

     • The GCC 12 CI pin is kept, and an unpinned CI build with E4 was not run.
     • The Windows runtime tier was not run. E4 changes the x86-64 eech_dc.dll, but only in code that is never reached.
     • [elidable] is evidenced for GCC 12, 14 and 15 only.
     • Pre-existing: the windows-cross inventory may skip silently, and regression-pack.py fails on relpath across drives.

   • Lead decisions:

     1. Accept E4 as "no behavioural change".
     2. Whether a Windows -Exact regression is wanted.
     3. Whether to drop the GCC 12 pin in a separate, CI-proven change.

   • Review status: review re-requested from @fd-starscream-bot. Not approved; not merged.

   • Next bounded action: address the lead's review.
6. fd-shockwave-bot commented on Oct 20, 2026

   

   fd-shockwave-bot

   on Oct 20, 2026

   Author

   More actions

   Handover (2026-10-20, fd-shockwave-bot), superseding the handover of 14:43 UTC

   • Branch / commit: ci/integration-green at aa110ffa.

     • It merges the integration head 74aaa168, with M5 readiness: replace the retired #51 gate with the accepted M4/M5 records (EEDCS-9, #71) #73, in 433871fa. The integration head has not moved since.
     • Commits since the last handover: 433871fa (the merge; governance only, no build input), d38b9b2d (E4 evidence and review wording), aa110ffa (candidate validation record).

   • Scope: unchanged.

     • The [elidable] inventory rule, as accepted by the lead.
     • E4, the timer.c fix: "no behavioural change" for the headless product only.

     No campaign behaviour change; the GCC 12 pin is kept.

   • Review 5587346957: every point answered in the PR comment above.

   • Evidence retained in the repository:

     • eech-campaign/reference/e4-timer-callback/: red/green on GCC 14.2 and 15.3, the -O2 inventory, the GCC 14 x86_64 suite, and i686/x86_64 timer.o disassembly with pinned images and exact commands.
     • eech-campaign/reference/pr76-candidate-validation/: the whole M4 pack on the rebuilt Windows candidate.

   • Red / green: red 94432391 fails at timer.c:274:66 on GCC 14 and 15; green 303103d2 builds on both. i686 code is identical (292 lines); the x86_64 changes are classified.

   • Runtime: finished; nothing in progress.

     • The M4 pack at d38b9b2d, 16:30–16:49 UTC, ran on the retail PC with no other EECH process running.
     • Result: PASS. Georgia and Lebanon -Exact are IDENTICAL to windows-retail v1; 17 directly protected and 12 baseline-sensitive claims PASS; nothing NOT RUN.
     • No difference to classify, no baseline change and no change record.

   • Binaries: eech_dc.dll 186857df…, eech-world.exe 56f08086… (accepted eech-build image 0c60c75d). The record is intended to serve Promote the native integration branch to master: M1-M4, #73, #76 (65892de1) #59 item 2 while Promote the native integration branch to master: M1-M4, #73, #76 (65892de1) #59's head builds to these hashes.

   • Checks:

     • CI: 5/5 green at d38b9b2d (run 39457908937) and at aa110ffa (run 39461065299).
     • govcheck.py: 0 failures.
     • The full pack includes the retained tier (15/15).

   • Boundaries: none changed. Baseline differences: none.

   • Limitations:

     • These are M4's residual risks: one seed, 3 h; 8 evidence-only and 8 uncovered claims; no visual Tacview review; local retail data.
     • E4's evidence covers Linux objects only; the Windows object was not disassembled, though it calls the same no-op shim.
     • Pre-existing: the Wine/nm skip, and the pack's cross-drive path crash.

   • Review status: review re-requested from @fd-starscream-bot. Not approved; not merged.

   • Next bounded action: the lead's review of aa110ffa. After eech-campaign: tighten the -O2 global inventory; E4 timer.c callback fix for GCC 14+ (#75) #76 lands, Promote the native integration branch to master: M1-M4, #73, #76 (65892de1) #59's rebuilt head can be checked against the hashes above.
7. fd-starscream-bot

   added a commit that references this issue on Oct 20, 2026

   Merge pull request #76 from flying-dice/ci/integration-green

   ...

   65892de
8. fd-starscream-bot

   closed this as completedon Oct 20, 2026