feat(parser): content-hash diff-summary disk cache (#845)

[gfargo]

Summary

PR 5 of the #845 diff-condensing pipeline sprint. Adds a per-repo, content-hash-keyed disk cache so re-runs of coco commit / changelog / recap / review skip the LLM entirely for files whose diffs haven't changed since the prior run. Cache writes are best-effort and gated by an opt-out env var.

What changed

• src/lib/parsers/default/utils/diffSummaryCache.ts — new module. JSON-on-disk cache with sha256 keys, 500-entry hard cap, LRU eviction, schema-versioned envelope. 19 unit tests cover key derivation, path resolution, round-trip, repo isolation, eviction, touch, clear, corrupt-file robustness.
• src/lib/langchain/chains/summarize/prompt.ts — exports SUMMARIZE_PROMPT_HASH (sha256 of the template, first 16 chars). Cache keys mix this in so any prompt edit invalidates prior entries automatically.
• summarizeFileDiff (in summarizeLargeFiles.ts) and summarizeDirectoryDiff (in summarizeDiffs.ts) — check the cache after the trivial-shape skip, return synthetic results on hit, write on success.
• src/commands/cache/ — new coco cache <subcommand> command. clear removes the file; info reports entry count, on-disk size, total summary tokens, last-saved timestamp.
• bin/benchmark.ts — new --repeat flag runs each fixture cold + warm so PRs can demonstrate cache wins. New --no-cache flag for reproducing pre-PR-5 numbers.
• .bench/baseline.json — refreshed to capture both cold and warm passes.

Cache key

`sha256(diff + '\x1f' + model + '\x1f' + promptHash)`

The model and prompt hash are mixed into the key, so:

• Switching providers/models invalidates entries (different models produce different summaries).
• Editing SUMMARIZE_PROMPT invalidates entries (different prompt → different output, no manual bumps required).
• Same diff text on the same model + prompt → cache hit, instant return.

Storage

• Path: `$XDG_CACHE_HOME/coco/diff-summaries/summaries.<sha1(repoPath, 16-char)>.json`
• Per-repo isolation (no cross-repo pollution; mirrors the boot cache from feat(log-tui): per-repo disk cache of the last commit-log fetch (#808) #828).
• 500 entries hard cap, LRU eviction on overflow (oldest `lastAccessedAt` evicted first).
• Best-effort: read failures fall back to "no cache" (LLM runs as before); write failures are swallowed silently. The cache is never load-bearing.

Opt-out

`COCO_NO_CACHE=1` disables both reads and writes for the entire run. Useful for benchmarking, debugging unexpected output, or forcing a full regeneration.

Bench numbers

Bench with `--repeat` against all fixtures (cold pass clears the cache first; warm pass runs again with full cache):

Fixture	Cold (ms)	Warm (ms)	Reduction
tiny	1	0	(early-exit, no cache touched)
medium	8506	4	99.95%
large	14417	6	99.96%
feature-add	13064	4	99.97%
refactor	51116	6	99.99%
initial-commit	16967	7	99.96%
docs-update	24384	2	99.99%
dep-bump	0	0	(early-exit, no cache touched)
monorepo	88958	26	99.97%

The cold numbers match PR 4's baseline (the cache miss path is identical to the prior code path). The win is the warm column: same fixture, unchanged inputs, second run is essentially free.

The realistic case for users is "edit one file → re-commit": cache hit rate ≥ N-1/N where N is the file count, so wall-clock collapses to one LLM call instead of N.

Test plan

• npm run test (5134 passing, 0 regressions; +24 new)
• npm run lint
• npm run build — confirmed CLI builds cleanly
• npm run build:schema — schema unchanged
• node dist/index.js cache --help — command surfaces correctly
• node dist/index.js cache info — shows real entries from a recent bench run (80 entries, 15.1 KB)
• npm run bench -- --repeat — confirms warm wins across all fixtures