ci-fast: vendor-hygiene compiler_minimal lint runs 1h+ — full fast gate not completable (#2265 fallout)

[PurHur]

Summary

ci-fast.sh cannot complete in reasonable time: the bootstrap vendor-hygiene PHPUnit stage runs bin/compile.php -l test/selfhost/compiler_minimal/main.php (112-require mega-bundle compiled as one Script), which runs 1h16m+ in the pinned container (killed, no completion observed; ~70min also with the #16394 opt-in frontend paths — the cost is the compile pass, not the frontend).

Because ci-fast dies (or is abandoned) before later stages, downstream gates rot silently: docs/bootstrap-inventory-triage-top50.json sat at a 2721-file snapshot vs 4038 live for months (refreshed in #16394 — now 0.5s warm via the lint cache), and this week three AOT-breaking merges sailed past the (unreached) AOT smoke gate.

Expected

Full ci-fast completable in well under an hour so agents actually run it to the end.

Candidates

• Compiler: producer-expr index + comment-interval cache for big-file compiles (#16077) #16412 (producer-expr index) directly attacks the compile-pass cost — findCfgProducerExprForOperand is quadratic in seen-blocks, and the mega-bundle maximizes it. Re-measure this stage after it lands.
• Cache the compiler_minimal lint result keyed by bundle content hash (same pattern as build/lint-cache).
• Or demote the mega-bundle lint to ci-local / a dedicated slow lane.

Done when

• time php bin/compile.php -l test/selfhost/compiler_minimal/main.php < 5 min in the pinned container, or the stage is cached/moved
• ./script/ci-fast.sh completes end-to-end on a clean master checkout

🤖 Generated with Claude Code

[PurHur]

Measurements with #16412 now merged (pinned container, compiler_minimal 112-require bundle via bin/compile.php -l):

• legacy paths, default 1536M: never completes (killed at 1h16m and 55m)
• producer index, default 1536M: OOM at 15m33 — the index retains O(exprs) on a mega-Script
• producer index, PHP_COMPILER_MEMORY_LIMIT=8G: 21m31, rc=0 — first observed completion

Remaining work for this issue: raise the memory limit for the vendor-hygiene bundle lint invocations (LlvmToolchain::envPrefix or the tests in BootstrapSelfhostBundleTest/BootstrapSelfhostLinkTest), and/or cache the bundle lint keyed by content hash (build/lint-cache pattern from #16394), and/or trim index memory (plain-array keying is unsafe due to spl_object_id reuse; a bounded index or per-function reset would be the safe cut). 21m is finishable but still the long pole of a full ci-fast.

[PurHur]

Status after the structural fixes (#16394 #16412 #16568 #16572 #16587 #16684 + spine syncs):

All gate layers ahead of the phpunit battery are now green and self-healing — two frozen-tree runs cleared capability/inventory/triage/wave3/AOT-smoke/vendor-hygiene in ~10 min (vs never before). What still prevents an end-to-end green run is test-suite content, not gate infrastructure:

1. libLLVM-9 segfault at a fixed offset in JIT compile subprocesses ~50% through the battery — detailed on Compiler SIGSEGV: ++/-- lowering crashes on null locals and under the FunctionStatic phpunit filter #16565 (both full runs died in the same region; three identical kernel traces).
2. for/while loop VM test effectively never finishes (VM: 1M-iteration loops exit 255 silently after ~30 s — benchmarks/ compute suite cannot run on the VM (bin/vm.php) #15906 class — burned 41 CPU-min before being reaped; needs a per-test timeout or smaller iteration count).
3. vm.php children left SIGSTOPped hang the suite at 0% CPU (phpunit battery: bin/vm.php children left in T (stopped) state — suite hangs until SIGCONT #16657; babysitter workaround documented there).
4. require-redeclare fatal killing the whole process at 49% — fixed in Test: require_once bootstrap-inventory-lint-sync-lib in both triage tests #16684.

Run wall-times for the record: run 5 died at 49% after 4h45m (fatal #4); run 6 died ~50% after 5h32m (segfault cluster + hang). Suggest keeping this issue for the battery-side items and closing once one full run exits with an honest pass/fail summary.

[PurHur]

claim: cursor-agent-lane-c — starting this run

[PurHur]

claim released: ci-fast currently fails earlier on spine coverage sync (missing new lib/JIT builtin units). Pivoting to fix that red gate first.

[PurHur]

claim: cursor-agent-lane-c — fixing spine coverage sync for StringHtmlspecialcharsDecodeLlvm.php (ci-fast blocker)

[PurHur]

Merged via #17595.

Fix: added StringHtmlspecialcharsDecodeLlvm.php to self-host spine + regenerated bootstrap inventory.

Verification:

• php script/check-selfhost-spine-coverage-sync.php → OK (4313/4313)
• php script/bootstrap-inventory.php --check → OK 4313/4313
• make bootstrap-selfhost-vm-driver-execute-probe → OK