fix: make gen-context recency boost deterministic (#440)

[manojmallick]

Problem

gen-context output was non-byte-stable run-to-run for ~5 large repos (fastapi, laravel, serilog, vapor, svelte), making the retrieval benchmark's hit@5 non-reproducible (86.7 → 85.6 → 87.8% over three unchanged runs). This is the residual tracked in #440 after the first pass (#441).

Root cause

The token-budget recency boost stamped mtime = Date.now() on every recently-committed file. On repos where nearly every file is "recently changed" (svelte: 8743 recent files), consecutive files often landed on the same millisecond. Which files shared a millisecond — and thus fell through to the filePath tie-break instead of sorting by a distinct mtime — shifted run to run, swapping which equal-priority files survived at the budget cutoff.

The Date.now() value wasn't just a boost: it encoded the (alphabetical) walk order as a monotonic mtime that bestFirst (mtime desc) relied on. The nondeterminism was only the millisecond collisions.

Fix

Replace Date.now() with a deterministic monotonic counter (nextRecentMtime) that reproduces the exact same processing-order ranking without collisions. Recent files stay ahead of non-recent ones and keep their walk-order ordering among themselves.

• An earlier attempt collapsing all recent files to one sentinel value regressed hit@5 to 63.3% (signalQuality dominated). The counter preserves the intended ordering.

Results

• All 43 benchmark repos: byte-identical across two clean runs (excluding the Updated: timestamp comment).
• Retrieval benchmark: identical per-repo results and 87.8% hit@5 across runs — quality unchanged, now reproducible. The published headline (87.8%) is no longer stale.

Acceptance criteria

• gen-context output byte-identical (excluding Updated:) across two clean runs for all benchmark repos
• Regression guard runs gen-context twice on a fixture and asserts byte-equality (test/integration/gen-context-determinism.test.js) — verified it fails on the old Date.now() behaviour
• Retrieval benchmark re-baselined to a single reproducible headline: 87.8% (weighted = unweighted over 18 repos)
• (Optional) omittable Updated: timestamp — deferred; touches 4 bundled-adapter sites + their source copies, disproportionate for an optional item. Guard already handles it by excluding the line.

Tests

• New: test/integration/gen-context-determinism.test.js (2 assertions)
• Full integration suite: 105 passed, 0 failed
• version.json derived test count 110 → 111

Supply-chain gate

Local audit PASS — no shell spawns · no install scripts · main exports API · no fingerprinting. Change is limited to an mtime assignment in gen-context.js.

🤖 Generated with Claude Code