feat(parser): skip pure-shape diffs in pre-process pass (#845)

[gfargo]

PR 2 of the #845 sprint. Headline: initial-commit-shaped repos drop from 60 s / 30 LLM calls to 10 s / 6 calls — an 84% wall-clock cut, no quality loss.

Why

Pure additions / deletions / renames-with-no-edit / binary file changes have no information content beyond the diff's shape. The LLM was getting a multi-thousand-token raw diff for + const foo = 1\n+ const bar = 2\n... and producing "Added foo.ts with N lines" — a string we can templated for free. New trivialDiff.ts helper detects those shapes from the hunk body and returns a deterministic summary; summarizeFileDiff short-circuits and skips the LLM entirely.

Detection rules

Cheap on purpose (runs per file in pre-process):

Shape	Detection	Templated summary
binary	Binary files X and Y differ header	Updated binary file \X`.`
rename	rename from + rename to headers AND no +/- body lines	Renamed \old` → `new`.`
addition	All body lines start with + (≥1 such line)	Added \file` (N lines).`
deletion	All body lines start with - (≥1 such line)	Removed \file` (N lines).`
modification	mixed +/- or rename-with-edit	undefined → LLM path stays

Headers (diff --git, index, ---, +++, @@, new file mode, deleted file mode, similarity index, rename from, rename to, Binary files) are ignored when classifying so the metadata --- /dev/null doesn't fool the deletion detector.

Bench (post-PR-1 baseline → PR 2)

fixture	calls before	calls after	Δ calls	wall before	wall after	Δ wall
tiny	0	0	0	1 ms	1 ms	0 ms
medium	19	6	-13 (-68%)	29,267 ms	6,906 ms	-22,361 ms (-76%)
large	30	6	-24 (-80%)	59,992 ms	9,749 ms	-50,243 ms (-84%)
feature-add	11	4	-7 (-64%)	19,591 ms	5,640 ms	-13,951 ms (-71%)
refactor	20	20	0	41,340 ms	41,347 ms	+7 ms
initial-commit	30	6	-24 (-80%)	60,034 ms	9,818 ms	-50,216 ms (-84%)
docs-update	7	7	0	18,563 ms	18,564 ms	+1 ms
dep-bump	0	0	0	0 ms	0 ms	0 ms

Reading the wins

Initial-commit shape (the user's #845 pain point) collapses 84%. A fresh git init with 50 source files used to take a minute of LLM work; now it takes ~10 seconds because all the pure-add files take the templated path.

Modification-heavy fixtures are unchanged, as expected — refactor (mixed +/- lines) and docs-update (markdown edits) still need real LLM work. They're targets for PR 4 (continuous-queue waves) and PR 6 (per-type prompts) respectively.

Why not 0 calls on the pure-add fixtures? Pre-process only runs for files larger than maxFileTokens (1024). Smaller pure-add files keep their raw diffs in the directory aggregate; the wave consolidation then summarizes those directories. Extending skip-trivial to wave consolidation (when all files in a directory are trivial) would eliminate those 6 remaining calls — possible follow-up PR if the marginal win is worth it.

Test plan

• npm run lint
• npm run test:jest (1265 tests pass — 15 new in trivialDiff.test.ts covering the four trivial shapes, headers-vs-body classification, line-count templating, singular wording for 1-line edits, and the rename-with-edit fall-through that MUST NOT classify as trivial)
• npm run build
• npm run test:cli
• npm run bench — numbers above

Plan reference

PR 2 of the #845 sprint. PR 3 (raise concurrency cap + adaptive backoff) is next, then PR 4 (continuous-queue waves), then PR 5 (disk cache).