perf(#3182): linear JS/TS func_start start anchor on blank runs (nx wall -73%)

[squid-protocol]

Part of #3182. Next fleet hotspot after #3317: typescript::Cartography_Mode_B_Braces (539s fleet CPU in v2.9.1), led by nx at 90 s/MLOC versus ~10 for grafana/sentry.

Root cause

The assignment branch of JS/TS func_start began (?:^|(?<=[^<>(,\s]))[ \t\n]* under re.M. _build_brace_safe_stream blanks a multi-line template literal to same-length whitespace, and the bare ^ fired on every line of the blank run. Each attempt scanned to the run's end before failing, which is quadratic: 4× per doubling, about 5.5s for 4,000 blank lines. nx's lock-file/__fixtures__/*.ts files (one 16k-line template each) cost 13–60s apiece, and the worst hit the 60s fuse. A sweep of every language's rules found this shape only in JS and TS func_start at fleet-relevant cost.

Fix

Inside a whitespace run, \b can only hold at the run's end, so every start in one run gives the same result. ^ is now taken only:

• on a non-blank line, or
• on a blank run's first line start (the previous line ends in a non-space), or
• at the file start.

These anchors are mutually exclusive. An earlier non-exclusive draft re-ran the failed continuation through a second anchor and was slower on a 1.3MB single-line file.

Evidence

	main	this PR	A-vs-A control
nx wall	70.8s	19.4s (−72.6%)	−0.1%
nx 5_Optical_Detector	320.5s	70.5s (−78.0%)	+0.9%
nx Cartography_Mode_B	90.5s	21.1s	noisy (±16%)
vscode Cartography_Mode_B	65.3s	62.4s (−4.3%)	0.0%
vscode wall	44.5s	43.9s	−1.2% (noise)

• audit.json and sarif are identical on both repos. On nx, every function_data row (name, start_line, loc, complexity, args) is also identical.
• Match-set parity across 58 JS/TS repos (~245k files, raw and brace-safe streams): identical except 21 matches in 20 files. In each of those, the OLD match started on a blank line after a line ending in <>(, plus trailing blanked text (a string or comment). The new match starts on the declaration's own line, with the same end and the same groups. Every one was machine-checked for that shape. This is a small line-attribution improvement: the old start sat 1–7 lines early.
• Golden crucible (full leg) passes unchanged. Locally, main and this branch both fail it with a cold .gitgalaxy dependency cache and pass warm, so that flake is pre-existing and environmental.
• ruff, mypy and dead-key audits are clean. The JS/TS and extraction suites pass (8,574).

New test tests/extraction/languages/test_func_start_blank_run_3182.py:

• a ReDoS guard (old anchor 5.5s, new 0.03s on the payload);
• an old-versus-new equivalence sweep over prefix × whitespace × suffix combinations, which allows only the documented shift.

Not in scope

• C# residue (runtime's TCFullEndElement.cs / MathF.cs, ~20s each): the stacked-[InlineData] bracket ambiguity, already tracked as perf: csharp func_start exponential backtracking on adjacent [...] (still fuses a roslyn file → 0 functions) #3320 (waits on the py3.11 floor, Raise minimum Python from 3.9 to 3.11 #3322).
• TypeScript/tests/cases/fourslash/codeFixClassImplementInterfaceNoTruncationProperties.ts (45s on main and on this branch). A 1.3MB blanked single line where branch 1's {0,200} annotation loop × lookahead whitespace scan is O(200·run). It's one file in the fleet, so I left it as-is.

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab