docs: raise the per-PR diff target to ~1200 hand-written lines

[thecodedrift]

Raises the per-PR diff target in CLAUDE.md from ~300 lines to ~1200, and excludes generated files from the count.

Why 300 did not hold

The rule spent its own budget. It counted tests toward the total while forbidding you to split them from the code they cover, and changes here are majority test code. #190 was roughly two thirds tests; #182 was 181 of 308 hand-written insertions.

Generated files inflated the count with nothing to review. #182 measured 515 lines, of which 346 was pnpm-lock.yaml and a regenerated schema. Against 300 that reads as a 1.7x overrun; against what a human actually reviews it is comfortably inside. The old rule did not say to discount them, so every judgement call began with an argument about what counts.

Prose changes have no natural 300-line seam. Agent recipes, CLAUDE.md, and OpenSpec artifacts routinely move several hundred lines in one coherent edit.

Why 1200 rather than 1000

The threshold has to be comfortable for a change carrying task changes, code, and tests together, which is the normal shape here. OpenSpec work adds proposal, design, spec deltas and tasks on top of the implementation, and the guidance elsewhere forbids splitting tests from their code. A number that only fits the code half pushes people to split along the exact seam the rest of the rules tell them not to cut.

Measured against recent work, 1200 sits above the coherent units and below the point where a diff stops being readable in one sitting:

PR	hand-written	at 1200
#182	308	comfortable
#187	~308	comfortable
#190	~1387	slightly over, correctly

#190 is the useful case: it really was four independently-safe units and could have been a stack, so a signal there is right. At 300 it was a 4.6x overrun, which is no signal at all.

One correction to the original rationale

The failure that motivated 300 was depth, not size. #103 reached "ready for review" as a ~93-file change never linted, typechecked, or tested, because a branches: filter silently stopped matching seven hops from main. That is a CI-coverage bug, separately fixed, and a 300-line target would not have caught it.

Note on the wording

The old sentence "a 900-line PR does not get reviewed, it gets approved" is kept in substance but re-anchored. At a 1200-line ceiling, 900 sits under the target, so naming it as the cautionary example would contradict the rule in the same breath. It now reads "a PR far past that".

Not changed

The preference for stacking, the merge-down rules, and changeset placement all stand. Only the size target and the generated-file exclusion move.

No changeset

CLAUDE.md is repo-root contributor guidance. It is not in packages/cli and ships in no artifact, so there is no user-facing behavior to describe. changeset.yml warns rather than fails.

Fixes #191