[squid-protocol]

Fixes #3913. Follow-up for the per-process import cost: #3914.

What was wrong

The Full Suite Gate's Windows jobs had reached their 30-minute limit. On #3901's own run, Windows 3.9 finished in 29 min 58 s, and on main's run 36335845364 both Windows jobs were cancelled at 98%.

Almost none of that time was testing:

• About 150 test fixtures scan a 3–6 file estate. Every scan started cpu_count - 1 worker processes, whatever its file count.
• Under the spawn start method, which Windows and macOS use, each worker re-imports GitGalaxy. That takes about 3 s, mostly compiling the language registry.
• Meanwhile xdist was running 4 such tests at once on a 4-core runner.

A one-file scan measured on Linux:

Start method	Wall time
fork	3.5 s
spawn	9.4 s

The extraction itself took 0.06 s.

Changes

• galaxyscope: no pool for a small scan.
  • It never starts more workers than files.
  • A scan of 16 files or fewer, or one run with GALAXYSCOPE_MAX_WORKERS=1, runs the same worker code in-process through _InlineExecutor.
  • The in-process path keeps what the process boundary gave:
    • the worker's arguments are copies;
    • the root log level and the SIGALRM ReDoS-fuse handler are restored;
    • exceptions arrive through a Future;
    • it runs on the main thread only.
  • A one-file scan under spawn drops from 9.4 s to 3.4 s.
• tests/conftest.py: under xdist, GALAXYSCOPE_MAX_WORKERS=1, because xdist already puts a test on every core.
• full-suite-gate.yml: timeout-minutes 30 → 45, as headroom only.

A real engine bug the new equivalence test found

detector.py decided whether to record call qualifiers (#3329) with invocation_pattern is p for p in QUALIFIED_CALLS_OUT_PATTERNS.

• A spawned worker receives its rules pickled, so each pattern arrives as a new object and the is check never matched.
• Result: every Windows and macOS scan silently dropped calls_out_qualifiers. On the test estate, 20 of 21 function rows differed from Linux.
• The golden masters run only on Linux, so nothing caught it.
• The family is now matched by the pattern's text.
• A forced-spawn scan now equals fork, which equals in-process, and a Linux test forces a spawned pool so this stays caught.

Two timing-sensitive tests, root-caused

• test_production_regex_redos_immunity called any 0.25 s gap in its status queue a hang.
  • A busy runner makes that gap by itself. It blamed a different innocent regex each run: dart::func_start on main's macOS, then cpp::func_start here.
  • The same rule also passed the test having checked nothing, whenever no worker had started within 0.25 s.
  • Now a regex is flagged after running for 10 s, and a stalled or incomplete run fails.
  • A new test proves the check still catches ^(a+)+$.
• test_a_redirect_is_refused_so_the_key_is_not_forwarded answered without reading the POST body.
  • On Windows, closing a socket with unread data sends a reset, which can arrive before the 302 (WinError 10053).
  • The handler now reads the body first.

Measured (Full Suite Gate, job wall time)

Job	Before (run 36340546823, #3901)	After (run 36346298642)
windows-latest 3.9	29 m 58 s (pytest 1648 s)	16 m 31 s
windows-latest 3.12	17 m 26 s	11 m 54 s (pytest 712 s)
macos-latest 3.9	27 m 44 s	9 m 20 s
macos-latest 3.12	19 m 38 s	8 m 47 s
ubuntu-latest (4 jobs)	5–10 m	5–10 m (unchanged)

The slowest job went from about 30 minutes to 16.5, so the 45-minute limit leaves room. Before this PR, the Windows 3.9 --durations list was led by the gauntlet oracle test at 318 s and about 20 mainframe fixtures at 38–47 s each. Now no mainframe fixture exceeds 23 s on Windows. Locally (4-core Linux, -n auto) the suite runs in 256 s, where the pre-push gates took 919 s earlier today.

Verification

• Golden masters are byte-identical in both modes. I checked twice:

  • with the whole crucible corpus forced through the in-process path (GALAXYSCOPE_MAX_WORKERS=1);
  • after the detector fix.

• audit_check.py: all clear.

• Full suite locally: 12,666 passed.

• New tests (tests/core_engine/test_inline_extraction.py):

  • in-process rows equal the pool's on a 21-file estate;
  • a spawned pool equals in-process on any OS;
  • a pickled rule is still recognised as the qualified-call family;
  • the executor's isolation guarantees hold.

  The spawn and pickling tests fail without the detector fix.

🤖 Generated with Claude Code

https://claude.ai/code/session_01X9PBcL8evG7CEzff1zE8v5

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab