full-test: the driver's proof skips the regen check its build already made; regen beside the shards; 4 Chrome shards

[thejackshelton]

What changed:

• full-test.yml:
  • regen input (boolean, default true). regen-chrome and regen-host run only when it is true or the run is not a dispatch, so a manual dispatch and a fulltest label run still check regen. The landing driver dispatches with regen=false (scripts/land-devices-ci.ts fullTestWorkflow).
  • Regen beside the shards: where the regen jobs still run, they need only resolve, not the Chrome shards. With the 4 Chrome shards, one of the 6 macOS jobs then waits for the free plan's 5.
  • 4 Chrome shards (was 3), with the summary expecting all four reports.
• scripts/test-durations.json: refreshed from runs 37399249352 and 37400820763 (231 files).

Why skipping regen is safe for the driver: buildPosition runs pnpm regen on every position it proves, and scripts/land.ts fails the build unless it exits 0. regen exits 0 only at a fixed point, a pass that changes nothing (MAX_PASSES 5, else "no fixed point"). The driver then runs the device lanes and pnpm regen again, to its fixed point, and commits the outcome. The regen-only check then confirms that commit holds only regen outputs. So the tree it proves is a fixed point by construction. The workflow's split regen --check would only re-prove that, with the same cache semantics and on the same toolchains.

Measured wall time of a land-test proof (dispatch to run completed):

run	regen	Chrome shards	wall time	note
37396636952 (before)	after the summary	3	30.5 min	summary at 11.9 min, regen-chrome to 30.5
37397958092 (before)	after the summary	3	24.7 min	summary at 11.4 min
37399249352 (before)	after the summary	3	~14+ min	regen-host still running at 13.7 min
37400820763 (after)	skipped	4	10.5 min	durations still master's (the run tests master's tree), so the shards were unbalanced: one at 10 min, three at 5.5–6.5
37402753106 (after, this branch's tree)	skipped	4	11.9 min	Chrome shards 5.1, 5.5, 6.0, 11.3 min; native 5.5–6.7; platform-free 3.3

• Target ≤ 12 min: met, but only just.
• The long pole now: Chrome shard 1, held by tailwind-sweep/sweep.test.ts alone, 300–414 s depending on what shares its runner. A shorter proof needs that file split (REGEN-SPEED/SLOW-SPLIT) or a fifth Chrome shard (the whole macOS cap).

What passed:

• pnpm typecheck passed.
• vitest run land-devices-ci test-shards land passed (112 tests).
• A new test checks:
  • the driver dispatches with regen=false;
  • the workflow declares the input with default true;
  • both regen jobs need only resolve and run only when the input allows.
• actionlint is clean. Both dispatch runs above passed: 231 files, summary success.

🤖 Generated with Claude Code