land: report a parallel build's ejections as they happen (queue 58: #91 and #198 failed with no FAILED line)

[thejackshelton]

Bug (queue 58, /tmp/land.log 01:23 and 01:51): in a parallel build (#209), each PR's failure went into a slot. prepareRound reported the slots (FAILED line, landing-failed label, PR comment) only after buildAll returned, which is after every position of the batch had built.

• In queue 58, INL1a 5/5 (C2): lowering inline boxes and <br>s, native runtime emit, 7 inline fixtures and lanes #91's preparation could not start, because of conflicts on master. That ended the speculation, so INL1a 5/5 (C2): lowering inline boxes and <br>s, native runtime emit, 7 inline fixtures and lanes #91 was built one by one and its merge failed (at 01:23). Then PNT1-radius: border-radius, its corner longhands and -webkit-border-radius drawn natively on iOS and Android (supersedes #80, #81) #198's merge on SVG-a1: inline <svg> (path, rect, circle; fill, stroke, stroke-width) on web, proven by a strict outline differential against Chrome; native refused as SVG-a2 #206's position failed (at 01:51). In both cases the log went straight to the next position with no FAILED line, and land.status showed "0 of 4 handled".
• The failures were not dropped, only delayed: they would be reported after ANIM-b slice 2: -webkit- transition/animation aliases, @-webkit-keyframes, and web frame fixtures for easings, direction and fill backwards #201's build (about 30 min later).
• A driver interrupted before then would never report them.

Fix:

• buildAll (and buildPositionsParallel) take a failed(index, error) callback, called the moment a position's PR is ejected: on assembly, a one-by-one fallback build, or after the speculation ended.
• prepareRound reports through its normal path at that moment, so the driver's round logs FAILED, labels and comments then, exactly as the one-by-one build does. It then skips that slot when it maps the results, so each failure is reported once.
• A builder's round collects the failures as before. A Fatal still stops the run.

What passed:

• pnpm typecheck passed.
• vitest run land-parallel land land-supervisor merge-train passed (148 tests).
• 2 new tests in packages/parity/test/land-parallel.test.ts. Both fail without the fix (verified by stashing the scripts/ change: 2 failed, 13 passed):
  • queue 58's shape: a preparation that could not start, then a one-by-one failure, which is reported before the next position builds, and once;
  • an ejection while assembling, reported before the positions above are built.
• The prepareRound slot test was changed to pass the new callback through (its signature changed).

🤖 Generated with Claude Code