Reduce MCP snapshot RSS

[justrach]

What changed

• Make snapshot writing stream uncached file contents from disk so large outline-only indexes still produce loadable snapshots.
• Cap fast snapshot restore content retention at 1000 files instead of loading every file body and freeing it afterward.
• Compact MCP snapshot-loaded sessions once ready, releasing raw contents when a valid disk word index is available.
• Make codedb snapshot persist a valid disk word index by streaming source files when the in-memory word index was skipped for a low-memory scan.
• Lazily load the disk word index for simple identifier codedb_search queries so exact/prefix searches stay fast without keeping raw contents resident.

Validation

• zig build test
• zig build
• Synthetic 3000-file MCP bench: contents cache drops from 3000 to 0; repeated identifier search median ~0.15ms after first lazy word-index load.
• Synthetic 2000-file large-content MCP bench: pre-search RSS drops from ~81MB to ~49MB; repeated identifier search median ~0.10ms after first lazy word-index load.

[github-actions]

Benchmark Regression Report

Thresholds: 10.00% and 50,000 ns absolute delta

NOISE means the percentage threshold was exceeded, but the absolute delta was too small to fail CI.

Tool	Base (ns)	Head (ns)	Delta	Abs Delta (ns)	Status
codedb_bundle	306889	327103	+6.59%	+20214	OK
codedb_changes	31519	32285	+2.43%	+766	OK
codedb_deps	5015	5035	+0.40%	+20	OK
codedb_edit	4356	4312	-1.01%	-44	OK
codedb_find	41304	46352	+12.22%	+5048	NOISE
codedb_hot	57903	55652	-3.89%	-2251	OK
codedb_outline	151594	162205	+7.00%	+10611	OK
codedb_read	56121	56053	-0.12%	-68	OK
codedb_search	143319	142765	-0.39%	-554	OK
codedb_snapshot	184308	184792	+0.26%	+484	OK
codedb_status	242661	233134	-3.93%	-9527	OK
codedb_symbol	39034	43390	+11.16%	+4356	NOISE
codedb_tree	54696	25676	-53.06%	-29020	OK
codedb_word	40530	39299	-3.04%	-1231	OK

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 94c9e0e92a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve rebuildable content when capping snapshot cache

This cap drops raw content for files after the first 1000 during fast snapshot restore, but Explorer.searchWord rebuilds the in-memory word index from explorer.contents when disk index loading fails. In repos larger than 1000 files where word.index is missing/unreadable (for example on a fresh machine with only codedb.snapshot), the rebuild becomes partial yet is treated as complete, so codedb_word and identifier-style search paths silently miss hits in uncached files.

Useful? React with 👍 / 👎.