Plan: steering php-compiler to a defensible v1.1.0 release

[PurHur]

php-compiler → release: steering plan

Written 2026-07-26. Every claim below was measured this session; sources are named so each can be
re-checked rather than believed.

Diagnosis

The project has a great deal of verification ceremony — gates, stamps, spine sync, generated-doc
checks — but the load-bearing checks do not run on the code that matters. So defects and
non-functional artifacts ship while the gates report green.

Evidence, all verified today:

finding	evidence
No CI on the compiler itself	.github/workflows/bootstrap-spine-gate.yml is the only workflow, and its pull_request trigger carries a paths: filter excluding lib/JIT/**, lib/VM/**, ext/**. A PR touching lowering gets zero checks, and gh pr view reports MERGEABLE/CLEAN — which means "no checks configured", not "checks passed".
Silent wrong output on ordinary PHP	24 of 43 mundane programs mismatched Zend. f($x+1, $x+2) printed 12 12; str_replace($p['from'], $p['to'], 'xy!') returned xy!. None caught by the compliance suite. Fixed in #23356/#23424.
The shipped compiler does not work	prelinked/bootstrap-gen0/bin-compile-aot fails parseAndCompile on every input including hello-world (#23468). Driver bytes are from 2026-06-15; 6441 lowering commits since; manifest restamped 272 times, provenance unverified-restamp.
A core cache was silently off for 20 days	Committed helper-runtime fingerprint 1771b935… vs live 6497e66d…. On mismatch every unit is skipped, so a two-line script recompiled ~257 units across 13 cores. Nothing noticed. Refreshed in #23457.
The AOT backend is wrong on its own benchmarks	mandelbrot renders all _ (#23471); Ack(3,8) segfaults (#23472). Both AOT-only — Zend and the VM are correct. These are the benchmarks the project quotes speedups from.
The test suites are too slow to run	VMTest ≈ 4 h serial; JITTest did not finish in 40 min. So they are not run per-change, and both carry large pre-existing failure sets (407 and 472).
Tooling could hang forever	script/bench.php had no timeout: bin/vm.php on Ack(3,10) ran 38 min at 100% CPU (Zend: 1.9 s) and blocked the whole suite. Fixed today.

The through-line: it is currently cheaper to make a gate green than to make the thing work.
Restamping satisfies freshness; a copy satisfies "native emit"; an empty filter satisfies CI.

Strategy

Make truth cheap and lies expensive. Concretely: every claim the repo makes about itself should be
produced by a check that fails when the claim stops being true, and no claim should be satisfiable
by restamping or copying.

Work in this order. Each phase exists because the next one is worthless without it.

---

Phase 1 — Make correctness observable (days, highest leverage)

Nothing else matters while defects can land unnoticed.

1. CI on lib/**, ext/**. Today: zero. This is the single highest-value change in the plan.
   Start with the cheap tier — php -l, unit suite, nikic preflight (18 s), differential sweep —
   so it fits a PR turnaround.
2. Differential sweep as a gate, VM and AOT (script/differential-sweep.sh, merged Test: differential sweep against Zend (#23354) #23444).
   It found all 24 argument defects and would have caught AOT: mandelbrot renders all '_' — escape condition wrong in native code (VM correct) #23471/AOT: Ack(3,n) binary segfaults (exit 139) — Zend and VM return correctly #23472. Extend cases toward the
   shapes users actually write.
3. Make the suites runnable. Sharding takes VMTest from ~4 h to ~20 min (24 shards, per-test
   TeamCity output so a stalled shard still reports). This session's harness works; land it as a
   script rather than leaving it in a scratchpad.
4. Quarantine flaky tests explicitly. hrtime_* (Regression: hrtime(true)/hrtime() pair non-monotonic — microtime refine walks ns backwards (re-#12225/#12279, ext/standard/hrtime.c) #23420), proc_get_status_basic,
   interface_abstract_static_call, dnf_return_type_error flip between runs and make regression
   comparison noisy — one appeared on the fixed list in one comparison and the regressed list in
   the next. Name them in a quarantine list; treat entry to that list as a bug to fix, not a
   dumping ground.

DECISION: compare suites by set difference of failing case names, never by count — neither
suite is green on master, so a count is meaningless.

Phase 2 — Stop shipping artifacts that do not work (weeks)

5. Fix the AOT correctness defects (AOT: mandelbrot renders all '_' — escape condition wrong in native code (VM correct) #23471 mandelbrot, AOT: Ack(3,n) binary segfaults (exit 139) — Zend and VM return correctly #23472 Ack segfault). Until these are
   closed, no speed claim about generated code is defensible.
6. Gate the gen-0 driver on function, not on a stamp (Bootstrap: committed gen-0 driver cannot compile anything — every input fails at parseAndCompile #23468): the committed driver must compile
   a script it has never seen, produce a runnable binary, and match Zend. A driver that starts and
   fails parseAndCompile currently satisfies every existing check.
7. Close the restamp loophole. Bootstrap: earn gen-0 provenance from the link, not the stamp (#22642) #22966 added build receipts, but BOOTSTRAP_GEN0_ALLOW_UNVERIFIED_STAMP
   is still exercised — two Trust: restamp gen-0 fingerprint commits landed today alone (Trust: restamp gen-0 fingerprint after lowering drift (#10533) #23429,
   Trust: restamp gen-0 fingerprint after lowering drift (#10533) #23438). Either the escape hatch requires a recorded reason and shows as degraded in
   release-readiness, or it goes.

OPEN: does any consumer actually need the gen-0 driver per-commit, or only per-release? If the
latter (likely), the per-commit freshness gate is the wrong shape entirely — see Phase 3.

Phase 3 — Break the artifact treadmill (structural, weeks)

The recurring failure is a committed artifact keyed on a fingerprint that changes several times a
day. Master merges ~4 commits/hour, ~56% touching lowering sources. Any artifact costing more than
~7 minutes to rebuild is stale on arrival, so it gets restamped instead.

8. Helper cache: per-unit dependency keys (Perf: helper-runtime cache key is too coarse — one lowering edit invalidates all 257 units #23458). Today one global core_fingerprint covers
   lib/JIT.php, lib/Runtime.php, … so one edit invalidates all 257 units. Key each unit on the
   closure it actually reached at emit time — the emitter already walks it.
9. Gen-0: split-TU build. The spine is ~96% partitionable (Bootstrap: measure how splittable the spine is before splitting it (#22642) #23018/Bootstrap: measure sub-splitting the oversized spine chunks — it does not work (#22642) #23033: only 3–4% of
   intra-spine class references cross a directory boundary). A single 6,519-file translation unit
   running ~4.6 h and OOMing becomes incremental, parallel and resumable — the model already exists
   in HelperRuntimeCache + emit-helper-runtime-object.php.
10. Treat gen-0 as a release artifact, not a per-commit invariant. Rebuild under a merge freeze,
    verify by function, tag, ship. Between releases, report staleness honestly
    (bootstrap-gen0-staleness.php already does) instead of restamping to green.

Phase 4 — Performance, with evidence (after 1–3)

11. Benchmarks that cannot hang — done today: per-run cap, separate build cap, n/a with a
    reason rather than a stall.
12. Only quote numbers traceable to a committed table where every runtime produced identical
    output. The current published AOT column was measured on binaries that are wrong or crash.
13. Then optimise, driven by profiles. The known hot spots are already measured: the php-cfg
    Simplifier quadratic path is 57% of gen-0 compile samples with a linear path sitting behind
    an opt-in env var, and helper-unit emit is O(units × transitive closure).

DECISION: no optimisation work before Phase 2 closes. Optimising a binary that renders
mandelbrot wrong is effort spent making a wrong answer arrive sooner.

---

Release criteria for v1.1.0

Falsifiable, and each one is a command someone can run:

1. script/differential-sweep.sh green on VM and AOT.
2. Committed gen-0 driver compiles an unseen script and matches Zend.
3. Compliance suites: no regressions vs the previous release by case name, flaky set quarantined
   and named.
4. benchmarks/README.md regenerated with no n/a in any column, every runtime output-verified.
5. No open defect of class "silent wrong output".

What not to do

• Do not optimise before Phase 2. Wrong-but-fast is worse than slow-but-right.
• Do not add gates a restamp or a copy can satisfy. The false-green family (Gen-1+ #1492: M4 ladder reports OK when gen-1 native emit FAILS and falls back to copying a prelinked sidecar (script/bootstrap-loop-gen1-link.sh) #21860: hardcoded
  emit_path=native, copy-fallback reported as native emit, substring match on
  native-prelinked-sidecar) is the template for what to avoid.
• Do not chase the gen-0 rebuild as a per-commit invariant. Arithmetic says it cannot be one.
• Do not trust CLEAN on a PR as evidence of anything until Phase 1 item 1 lands.

Sequencing note

Phase 1 items 1–3 are days of work and unblock everything else, because they turn "did I break
something?" from a 4-hour question into a 20-minute one. That single change is what makes the rest
of this plan executable at the project's actual merge velocity.

[PurHur]

Worker lane A — no eligible claim (2026-07-26)

Could not comment on #10533 (GitHub hard cap: 2500 comments / Commenting is disabled). Parking status here.

Gates: ./script/release-readiness.sh --json → user_release_ready: yes (inventory OK 6581/6581, spine sync, north-star5-fast, vm-driver-probe). Note: gen0_provenance still stale (driver 2026-06-15).

Lane A (n % 3 == 0, prefer language): no claimable issue.

• label:language / phase-3:language open: 0
• Remaining implementation-ready − stdlib Lane A tickets are meta (Strategic alignment: compiler excellence pillars (2026 H2) #10533/🗺️ ROADMAP: Self-host + web-capable PHP subset (living) #78), CI-only (CI: Default SESSIONS_WEB_DEPLOY_SMOKE_GATE=1 in ci-local after #1967 green #1974), or named-ext / ext/standard ports owned by STD Lib / Ext automations
• Open queue is dominated by stdlib-labeled Reflection/arity regressions — out of lane

Stopping per worker rules (do not steal other lanes). Lane A needs fresh language / compiler-core implementation-ready filings (or #10533 successor issue once comments are capped).

[PurHur]

Worker lane B — gate probe / empty eligible queue (2026-07-27)

Could not comment on #10533 (comments disabled at 2500+). Cross-posting status here.

TRY_FIRST

Gate	Status	Note
inventory --check	green	OK 6586/6586 (via release-readiness)
release-readiness --json	green (re-run)	user_release_ready: yes; north-star5-fast + vm-driver-probe ok
first readiness pass	flaky	vm-driver-probe once failed (lib-spine-smoke-link: failed to build compiled driver); re-run green
bootstrap-selfhost-link	green	Zend gen-0 fallback; committed gen-0 seed stale (#21855 fingerprints)
bootstrap-selfhost-helloworld	red	emit-helper cold link OOM @ 6GiB; Zend path free(): invalid pointer (134). Exclusive agent-harness-phpc-gen0-23468-r4 in flight — deferred
gen-0 provenance	stale	9e6b9c12a / 2026-06-15; 6545 lowering + 275 manifest rewrites

Lane B

• No unclaimed non-stdlib/non-ext lane-B IR issues.
• Language: Generator::rewind() after current()/key()/valid() throws — Zend allows at first yield (Zend/zend_generators.c) #23713 (Generator rewind) already claimed by another lane-B worker — not stolen.
• Stopped this run per lane discipline.

[PurHur]

Worker lane A — no eligible issue this run (2026-07-27)

Could not post on #10533 (GitHub: commenting disabled past 2500 comments).

TRY_FIRST

• docker info: OK
• php script/bootstrap-inventory.php --check: doc drift only (warnings 21289→21290) — not a link failure
• Prior maintainer probe (Regression: Generator::send() on unstarted value-yield returns first yield — Zend injects+advances (re-#18108, Zend/zend_generators.c) #23712): user_release_ready: yes, north-star5-fast OK
• Lane A Regression: Generator::send() on unstarted value-yield returns first yield — Zend injects+advances (re-#18108, Zend/zend_generators.c) #23712 already merged as Language: Generator::send() on unstarted value-yield matches Zend (#23712) #23717 (Generator::send value-yield) before this run could claim new work

Lane A scan (number % 3 == 0, implementation-ready, no open PR)

• Prefer phase-2:language: zero left (only Language: Generator::rewind() after current()/key()/valid() throws — Zend allows at first yield (Zend/zend_generators.c) #23713 lane B, Language: new Class(...) first-class callable under PROFILE=8.4 still compile-fatals (re-#9767, Zend/zend_compile.c) #23714/Language: UnhandledMatchError still leaks string subjects — not '...' (re-#23664, Zend/zend_exceptions.c) #23678 lane C remain)
• Remaining Lane A ready issues are owned by STD Lib / Ext automations (stdlib label or named ext/*) or are meta/CI (🗺️ ROADMAP: Self-host + web-capable PHP subset (living) #78, CI: Default SESSIONS_WEB_DEPLOY_SMOKE_GATE=1 in ci-local after #1967 green #1974)
• Did not steal other lanes

No claim / no PR this run.

[PurHur]

Worker lane A — no eligible claim (2026-07-27)

#10533 comments are locked (2500+), so parking status here.

Trust: release-readiness --json → user_release_ready: yes (inventory/spine/north-star5-fast/vm-driver-probe). gen-0 seed stale (known).

Disposition: Closed #23721 as not-planned — Zend/RFC reject class-const new; compiler already matches under PROFILE=8.4 (#21493).

Lane A empty: remaining open implementation-ready tickets in % 3 == 0 are STD Lib / Ext-owned / roadmap-meta; only other language IR tickets are lane C (#23720, #23678). No PR opened.

[PurHur]

Worker lane B — no eligible claim (2026-07-27)

Could not comment on #10533 (GitHub disables comments at 2500+). Status belongs there; mirroring here.

Pre-triage gates:

• bootstrap-inventory --check → OK 6586/6586
• release-readiness --json → user_release_ready: yes (inventory, spine, root-readme, north-star5-fast, vm-driver-probe ok)
• gen0_provenance: stale (informational)
• No in-flight PR/branch for this worker

Lane B (n % 3 == 1) scan: 34 implementation-ready open issues; none eligible — all owned by STD Lib (ext/standard Reflection/arity) or Ext automations (date/simplexml/spl/mbstring/ftp/fileinfo/pcre/curl/…) or large named-ext umbrellas (#3724 soap et al.).

Out-of-lane free language tickets (not claimed): #23720/#23678 (lane C), #23574/#23706 (lane A).

Stopped per lane rules.

[PurHur]

Worker lane A status (2026-07-27) — no eligible claim

Posting here because #10533 is at the GitHub 2500-comment cap (Commenting is disabled).

Trust gates: ./script/release-readiness.sh --json → user_release_ready: yes (inventory 6588/6588, spine sync, north-star5-fast, vm-driver-probe OK). gen-0 seed still stale (known).

Lane A sweep: no open implementation-ready issue that is (a) n % 3 == 0, (b) not meta/roadmap/parity-snapshot, (c) not owned by STD Lib / Ext / Runtime automations, (d) without an open PR / recent foreign claim:. Prefer-language found only #23824 (Lane B). Remaining Lane A queue is Reflection/named-arg stdlib or ext/* parity.

Stopped per lane rules — not claiming across lanes.

[PurHur]

Worker lane B — no eligible claim (2026-07-27)

Posted here because #10533 rejects new comments (>2500 cap; same as #1492).

Gates (this host):

• php script/bootstrap-inventory.php --check → OK 6588/6588
• ./script/release-readiness.sh --json → user_release_ready: yes (inventory, spine, readme-sync, north-star5-fast, vm-driver-probe all ok)
• First failing command: none
• Note only: gen0_provenance.status=stale (driver 2026-06-15 / 9e6b9c12a; 6601 lowering commits since)

Triage: lane B (n % 3 == 1) open implementation-ready queue is entirely owned elsewhere (stdlib / named ext/* / PECL). No language/bootstrap/IMPORTANT/area:jit|aot|compiler candidates. Did not steal. Open PR set only links #23483/#23480.

[PurHur]

Worker lane C — no eligible claim (2026-07-27)

Posting here because #10533 is at the GitHub 2500-comment cap (Commenting is disabled on #10533 / #1492).

Pre-triage gates (this host):

Gate	Status	Note
docker info	green
php script/bootstrap-inventory.php --check	green	OK 6588/6588
./script/release-readiness.sh --json	green	user_release_ready: yes (inventory, spine, readme-sync, north-star5-fast, vm-driver-probe all ok)
gen-0 provenance	warning only	stale seed since 2026-06-15 / 6615 lowering commits — honesty warning, not treated as red

Lane C scan (n % 3 == 2, open, implementation-ready):

• ~30 open implementation-ready issues in-lane; all are labeled stdlib and scoped to ext/standard Reflection/named-arg/null-coercion regressions, or named extensions (ext/intl, ext/mbstring, ext/date, ext/session, ext/opcache, …).
• Per worker rules: not claiming — owned by PHP Compiler STD Lib / Ext — DOM & XML / Ext — I/O & Crypto.
• Already claimed in-lane (last 48h): #23858 (claim: PHP Compiler STD Lib), #23264 (STD Lib claim).
• Non-stdlib in-lane leftovers are meta/epic only: #23474 (this plan), #23480 (sideload arch), #23483 (perf) — not single-repro implementation-ready tickets.
• Language ticket #23845 (final property) is lane B (n % 3 == 1) — not stealing.

Action: stop this run; no PR. Next lane-C implementer should re-scan for new phase-2:language / bootstrap tickets, or wait for STD Lib / Ext queues to clear so language backlog is filed again.

[PurHur]

Worker lane A — no eligible claim (2026-07-27) — #10533 comments locked

Could not post to #10533 (Commenting is disabled on issues with more than 2500 comments). Recording lane-A stop here.

Gates: inventory OK 6588/6588; release-readiness --json → user_release_ready: yes (north-star5-fast + vm-driver-probe ok). gen-0 provenance still stale (driver 2026-06-15 / 6615 lowering commits since).

Triage: sole open phase-2:language is #23845 (lane B, already claimed). All other lane-A implementation-ready tickets are stdlib/named ext/* (owned automations) or meta/CI. Stopping without stealing another lane.

[PurHur]

Worker lane C status (2026-07-27) — #10533 comments locked

Pillar 1: release-readiness --json → user_release_ready: yes (north-star5-fast + inventory 6590/6590 + VM driver probe ok). gen-0 seed still stale (known).

Merged in-flight:

• Test: cover script-scope pre-decrement and mixed ++/-- runs (#23840 follow-on) #23867 (g09 pre-dec differential coverage)
• Bench: document PHP_COMPILER_OPT_LEVEL, and correct my wrong Ack diagnosis (#23483) #23828 (OPT_LEVEL docs on Perf: generated code is 4-14x SLOWER than Zend on basic loops and calls #23483)

No new claim: every lane-C (n%3==2) implementation-ready open is stdlib/ext/*-owned by other automations. Only non-owned lane-C opens are #23483 / #23480 (not implementation-ready). Stopping per lane rules.

[PurHur]

Maintainer gate probe + parity filings (2026-07-27)

Could not comment on #10533 (comments disabled at 2500+). Status here.

TRY_FIRST

Gate	Status	Note
php script/bootstrap-inventory.php --check	green	OK 6588/6588
./script/release-readiness.sh --json	green	user_release_ready: yes (inventory, spine, root-readme, north-star5-fast, vm-driver-probe)
gen0_provenance	stale	driver 9e6b9c12a / 2026-06-15; 6621 lowering commits + 275 manifest rewrites since
first failing command	none (quick mode)	honest_compile left unknown (not run full metric)

Pillar 1 quick gates green → filed language/stdlib parity (not blocked on trust).

Part A — filed (implementation-ready)

• Regression: set-only / implicit-backed property hooks — Must not write to virtual property (re-#22452, Zend/zend_property_hooks.c) #23881 — set-only / implicit-backed property hooks → Must not write to virtual property (re-Regression: property set-only hook read fatals — Zend returns backing (re-#19163, Zend/zend_object_handlers.c) #22452)
• Language: attributes on file-scope constants parse-error under PROFILE=8.5 — Zend accepts (Zend/zend_language_parser.y) #23882 — file-/ns-scope constant attributes parse-error under PROFILE=8.5
• Stdlib: mb_trim/mb_ltrim/mb_rtrim invalid encoding — LogicException not ValueError (ext/mbstring/mbstring.c) #23883 — mb_trim* invalid encoding → LogicException not ValueError

Part B

• PR AOT: ReflectionProperty::isFinal() for plain final properties (#23845) #23874 (Regression: final plain property MODIFIER_FINAL ignored — child override compiles, ReflectionProperty::isFinal() false (re-#23683/#23824, Zend/zend_compile.c) #23845), Test: cover script-scope pre-decrement and mixed ++/-- runs (#23840 follow-on) #23867, Bench: document PHP_COMPILER_OPT_LEVEL, and correct my wrong Ack diagnosis (#23483) #23828 already merged before/during this run — no open merge candidates left in the short queue.
• Open queue still well stocked with implementation-ready (incl. today’s Regression: array_find/array_find_key/array_any/array_all advertise phantom $strict — Zend expects exactly 2 args (re-#6949, ext/standard/array.c) #23875–Regression: request_parse_body Reflection empty — Zend options; named options rejected (ext/standard/http.stub.php) #23878 Reflection/clone-with filings).

Not covered

• bootstrap-selfhost-link / helloworld (not re-run; prior sessions noted gen-0 stale + possible OOM)
• Full ci-local / differential sweep

[PurHur]

Maintainer gate probe + backlog stock (2026-07-27)

Could not comment on #10533 (2500+ comment cap). Cross-posting here.

TRY_FIRST

Gate	Status	Note
release-readiness --json	green	user_release_ready: yes; inventory OK 6590/6590; spine; north-star5-fast; vm-driver-probe
gen-0 provenance	stale	9e6b9c12a / 2026-06-15; 6625 lowering + 275 manifest rewrites
Open PRs	none	#23890 (Closes #23882) already merged

Filed implementation-ready (VM-verified today)

• Language: bare local as call argument inside function is undefined — echo/expr OK (Zend/zend_execute.c) #23893 Language: bare local as call argument inside function is undefined (VM+JIT) — lane A / compiler-core
• Regression: ReflectionConstant getFileName/getExtension*/getDocComment/inNamespace still missing on PROFILE=8.4 (re-#22662, ext/reflection) #23894 Regression: ReflectionConstant getFileName/… still missing on PROFILE=8.4 (re-Regression: ReflectionConstant getFileName/getExtension*/getDocComment/inNamespace still missing on PROFILE=8.4 (re-#22421, ext/reflection) #22662)
• Regression: array_first/array_last Reflection empty — Zend array; named array rejected (ext/standard/array.stub.php, PHP 8.5) #23895 Regression: array_first/array_last Reflection empty / named array: rejected (PROFILE=8.5)

No open worker PR to merge this run; IR queue remains stocked (~99).

[PurHur]

Worker lane B — TRY_FIRST / no eligible claim (2026-07-27)

Note: #10533 comments are disabled (2500+); posting here as the living release/steering tracker. Lane B (issue % 3 == 1).

TRY_FIRST gates

Gate	Status	Symptom
php script/bootstrap-inventory.php --check	green	OK 6588/6588
./script/release-readiness.sh --json	green	user_release_ready: yes; north-star5-fast ok; spine / readme / vm-driver-probe ok
gen0_provenance	stale (warning)	driver from 9e6b9c12a (2026-06-15); 6620 lowering commits / 275 manifest rewrites since
make bootstrap-selfhost-link	green (partial)	OK /compiler/build/selfhost via Zend gen-0 fallback; committed gen-0 refused stale (#21855)
make bootstrap-selfhost-helloworld (STRICT)	red	emit-helper OOM — see below

First failing command: make bootstrap-selfhost-helloworld → inventory compile_driver emit-helper link (#3032)

PHP Fatal error: Allowed memory size of 25769803776 bytes exhausted … in /compiler/lib/JIT.php on line 1197
emit helper link failed (exit 255, mode=inventory compile_driver (#3032))
BOOTSTRAP_M3_HELLOWORLD_STRICT=1 — refusing Zend fallback

Retried with PHP_COMPILER_DOCKER_MEM=32g + PHP_COMPILER_LLVM_MEMORY_LIMIT=24576M — still OOM at 24 GiB. Zend-fallback then hit Cannot redeclare PHPCfg\Parser::extractGlobalDeprecatedConstMarkerPayloadFromAttributes().

Lane B queue

No eligible claim: every open % 3 == 1 ticket is stdlib / named-ext/* (other automations), Parity backlog snapshot, or meta. Did not steal Lane A #23871 (sprintf/printf AOT regression).

Next (not started)

Honest path is gen-0 / emit-helper refresh under exclusive high-mem (not a restamp), or shrink/split the helloworld inventory emit bundle — STRICT helloworld no longer fits ≤24G when sidecars are stale.

Stopping without a PR (no in-lane issue; pillar-1 quick gate still green).

[PurHur]

Maintainer gate snapshot (2026-07-27) — #10533 comments locked

Tried software first via ./script/docker-exec.sh.

Gate	Result
bootstrap-inventory.php --check	OK 6590/6590
release-readiness --json	user_release_ready: **no**
First failing gate	vm-driver-probe — spine-smoke compiled driver build fail in readiness; standalone probe reports OK with stale sidecar stamp (want a7b15e8… / have 1f682d54…)
gen0_provenance	stale (driver 2026-06-15; 6k+ lowering commits since)
Open PRs	none this run

Filed implementation-ready parity (verified VM PROFILE=8.4): #23897 (final plain props), #23898 (new Foo->), #23899 (curl upkeep constants). ~101 ready issues in queue.