fix(index): bound searchPrefix allocation by max_results

[justrach]

Summary

• Add max_results: usize parameter to WordIndex.searchPrefix (src/index.zig:271) so the prefix-key scan stops once the cap is reached, instead of collecting every deduplicated hit and relying on the caller to truncate.
• Pre-size the result list and dedup map to the cap, switch the append to appendAssumeCapacity, and break out of both the inner hits loop and outer keyIterator once result.items.len >= max_results.
• Update the only caller, Tier 0.5 in Explorer.searchContent (src/explore.zig:1314), to pass the user-requested max_results.

Addresses Codex's P2 review on #293: on large indexes a broad 3-char prefix could allocate a huge hit set before the post-call truncation in searchContent, causing latency spikes / OOM despite a small caller limit.

Test plan

• zig build test — 388/388 pass (387 baseline + 1 new word-index: searchPrefix respects max_results cap).
• New test indexes 50 distinct files with the fooBar prefix and asserts searchPrefix("foobar", a, 5) returns at most 5 hits.
• Existing two searchPrefix tests updated to pass an explicit cap (32) and continue to pass.

🤖 Generated with Claude Code

[github-actions]

Benchmark Regression Report

Threshold: 10.00%

Tool	Base (ns)	Head (ns)	Delta	Status
codedb_bundle	519685	526987	+1.41%	OK
codedb_changes	90129	86340	-4.20%	OK
codedb_deps	44071	41834	-5.08%	OK
codedb_edit	17027	17271	+1.43%	OK
codedb_find	80138	80976	+1.05%	OK
codedb_hot	134605	130104	-3.34%	OK
codedb_outline	313358	313875	+0.16%	OK
codedb_read	133522	121252	-9.19%	OK
codedb_search	224262	221994	-1.01%	OK
codedb_snapshot	3749475	3756895	+0.20%	OK
codedb_status	230676	230695	+0.01%	OK
codedb_symbol	92570	91019	-1.68%	OK
codedb_tree	88555	88527	-0.03%	OK
codedb_word	120527	118252	-1.89%	OK

[chatgpt-codex-connector]

💡 Codex Review

codedb/src/watcher.zig

Lines 602 to 603 in 8f2fc19

	for (extractors[extractors_done..]) |*r| r.deinit(allocator);
	allocator.free(extractors);

Track initialized extractors before deferred cleanup

In buildTrigramsFromCache, the deferred cleanup iterates extractors[extractors_done..], but extractors_done is only incremented after successful post-join teardown, not after initialization. If std.Thread.spawn fails partway through the setup loop, this defer runs and calls deinit on uninitialized TriExtractResults entries, which can crash or corrupt memory during an otherwise recoverable error path.

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".