Warm rector daemon trips ClassReflection bug on test classes — real fix behind PR #272 band-aid

[fdaviddpt]

Background

PR #272 stopped a poisoned-cache symptom: mcp-rector-warm's long-lived daemon intermittently emits rector's own System error: "ClassReflection must be resolved for class XTest" on test classes, and that failure was being cached + replayed (2100 poisoned entries found).

PR #272 papers over it at two layers (drop System error: in the rector-mcp adapter; never cache non-deterministic engine failures). That's correct — an engine glitch is not a code finding — but it's a band-aid. The warm daemon still hits the reflection bug; we just stop surfacing/caching it.

Root cause (as far as proven)

• Plain rector CLI (fresh boot) reflects the same test file clean (exit 0).
• The warm in-process container (RectorRunner, reused across calls) intermittently fails to resolve a ClassReflection for the analyzed test class. It's warm-process-state dependent, not file dependent — cold-first and warm-repeat both passed in controlled runs; the live failure only reproduced from a deeply-warm daemon, then got frozen into the content-hashed cache.
• Most likely: BetterReflection in the warm process takes an 'already known' path for a class it can't resolve from rector's scoped source locator. Exact trigger not pinned — needs a deterministic repro inside the warm process (run rector --debug in the embedded container and capture the stack).

What a real fix looks like

Make the warm daemon's reflection scope match a cold CLI boot so test classes resolve — e.g. ensure the test autoload paths / reflection provider are warmed the same way, or reset the relevant reflection state per call. Lives in the mcp-rector-warm repo, not here.

Acceptance

• Deterministic repro of the warm-only ClassReflection failure.
• Warm daemon returns the same result as rector CLI for test classes (no System error:).
• Once fixed upstream, the adapter's System error: suppression can stay as defense-in-depth but is no longer load-bearing.

Ref: PR #272.

[fdaviddpt]

Resolved

Root cause found and fixed upstream in mcp-rector-warm #2 → released v0.4.0.

Mechanism (deterministic, exactly as suspected here): the warm daemon reused one Rector Application and reset nothing between calls. Rector's DynamicSourceLocatorProvider caches its AggregateSourceLocator for every non-PHPUnit run, so the second file in a warm session was analysed with a source locator that only knew the first file — and its own ClassReflection couldn't be built → System error: "ClassReflection must be resolved for class X". A fresh rector CLI never hit it (one process per call). RectorRunner::run() now resets every ResettableInterface service before each warm call — the same flush AbstractRectorTestCase does between fixtures.

Acceptance

• ✅ Deterministic repro. Feed real test files through one warm server: it fails at file feat(gl-mr): list conflicting files + hunks + resolve commands #2 — AStartTest.php (warms) → AresGlobalModuleTest.php → System error: "ClassReflection must be resolved for class AresGlobalModuleTest". Cold (fresh server, same file) is clean → warm-only. Harness committed as tools/repro-273.py; env-gated regression testWarmReflectionMatchesColdForSequence.
  • Note: a self-contained fixture can't reproduce it — in-process PHPUnit sets isPHPUnitRun()=true which disables the locator cache, and the trigger needs real framework base classes extended from outside the configured paths. That's why earlier controlled 2-call runs passed.
• ✅ Warm daemon == cold CLI for test classes. 300-file warm sweep: zero System errors after the fix; reverting reintroduces the failure at file feat: between op — extract function bodies and line slices #1.
• ✅ Suppression is now defense-in-depth, not load-bearing — #285 (merged).

Downstream

• mcp-rector-warm v0.4.0 released (Packagist).
• supertool docs(rector-mcp): mark System error suppression as defense-in-depth (#273) #285 merged (requires mcp-rector-warm ≥ 0.4.0).
• DVSI bumped ^0.2→^0.4 (MR !24633).

Cost note: ~0.4s/warm call, unchanged — bootstrap stays warm, only per-run reflection state is flushed.