perf: traversal micro-optimizations

[alyx]

Stacked on #2 (merge that first). Six low/medium audit findings in the traversal path, all semantics-preserving, plus two findings re-measured and deliberately skipped.

Changes (one commit each)

• detect: marker index — markers split into literal names (map lookup) vs true globs at engine construction; *.sln-style globs reduce to HasSuffix. Walk-side detection CPU drops ~15× on marker checks. The audit prototype's shared scratch slice was redesigned into a per-hit local allocation because it would have been a data race under perf: counter fast path, single-fd reads, parallel directory walk #2's parallel walk (race detector run on all touched packages).
• ignore: split paths once — each directory entry's rel path is split into segments once and fed to gitignore-stack, exclude/include-pattern, and smart-detect scope matching; gitignore source descriptions are no longer built when tracing is off.
• walker: per-entry overhead — path.Join/filepath.Join → plain concatenation (inputs already clean), names slice only built when smart detection is on, trace-arg boxing guarded behind a cached flag, seen map + mutex skipped for single-root/no-symlink runs.
• walker: lazy display paths — the display string is built only when something reads it (by-file detail, dedup, warnings, traces); aggregate runs skip ~1 string per counted file.
• walker: tracked-dir ancestor loop — breaks at the first already-present ancestor (was O(files × depth)).

Skipped with measurements

• darwin worker cap: after perf: counter fast path, single-fd reads, parallel directory walk #2's parallel walk, default jobs vs -jobs 8/-jobs 4 are indistinguishable in wall time — the win the audit measured predated the parallel walk. Not applied.
• GOGC tuning: GOGC=800 is now ~1% user CPU (below the 3% bar) — perf: counter fast path, single-fd reads, parallel directory walk #2's pooled reads removed the garbage that made it pay. Not applied.

Evidence

• User CPU: −18.7% on the 52k-file corpus (6 interleaved pairs, low variance), −7.3% on Linux 6.9; mallocs −25.8% on Linux 6.9 aggregate.
• Byte-identity vs perf/hot-path across: aggregate/by-file/dedup JSON, YAML, table, Linux 6.9, GOROOT/src, --tracked (plain and -vv), file/dot roots, multi-root overlap (duplicate-root trace intact), all filter flags, -v summary, full -vv on the 52k corpus. Only exception: -vvv per-file lines interleave nondeterministically on the base binary too (worker-emitted; sorted output identical).
• go test ./... green; -race green on walker/cli/detect/ignore; cross-builds for windows/linux clean.

🤖 Generated with Claude Code