search/nav: high trajectory variance; occasional non-convergent runaways (3-5x tokens) on large repos

[justrach]

Summary

The index file codedb.snapshot is written into the indexed project's root directory, where it shows up in git status and can be accidentally committed. It is large - in one real project the root codedb.snapshot is 22.8 MB.

Evidence

• A codedb.snapshot (22.8 MB) sits at the root of an actively-developed repo after normal codedb use.
• In an automated harness, a plain git add -A && git diff swept codedb.snapshot into the captured patch (it began with diff --git a/codedb.snapshot ... Binary files differ), corrupting the diff. Workaround required adding it to .git/info/exclude per checkout.

Impact

• Pollutes git status; easy to commit a multi-MB binary by accident.
• Breaks any tooling that diffs/snapshots the working tree.

Suggested direction

• Store the index outside the working tree (e.g. an XDG cache dir keyed by repo path), or
• Write it to a .codedb/ directory and auto-append codedb.snapshot/.codedb/ to the repo's .gitignore (or .git/info/exclude) on first index.

Found via an independent SWE-bench Lite token-efficiency benchmark: identical agent (`claude -p`, Sonnet 4.6) and tasks, only the tool surface differs - native Read/Grep/Edit vs codedb MCP tools. N=51 paired instances; full harness + data available.

[justrach]

Fixed in 227ea68: a per-session ConvergenceGovernor tracks recent nav call signatures and, when the same search/read recurs ≥3×, appends an in-band nudge steering the agent to a structural tool / direct read / refined query — a cheap interception for the 3–5× runaways. Never alters results; write/admin tools aren't governed. Merged to release/0.2.5825 and release/0.2.5826; PR #631 → main. Test: issue-624: convergence governor flags a repeated identical call.

Note: this is the interception layer; full empirical A/B (does the median hold while the tail caps?) still wants a run through your bench/swe-lite harness.