fix(build): turbo runner now actually reports per-task accounting (#830)

[Dave-London]

Summary

Closes #830.

mcp__pare-build__turbo was returning {success: true, totalTasks: 1, passed: 0, failed: 0, cached: 0, tasks: []} on every modern (turbo 2.x) run — claiming success while reporting zero per-task data.

Root cause

Turbo 2.x emits per-task status lines as <pkg>:<task>: cache hit/miss/bypass ... (colon separator, no trailing (duration)). The parser regex TURBO_TASK_RE was hard-coded to ^(.+?)#(\S+):\s+cache\s+(hit|miss|bypass)(?:,\s+\w+[^(]*)?\(([^)]+)\) — requiring # between package and task and a mandatory (duration) suffix. Every per-task line silently failed to match, so tasks: [] while totalTasks came from the Tasks: N successful, M total summary line. Failure parsing was broken in the same way.

The 1.9s duration mentioned in the issue is turbo actually running — turbo just refused to print extra accounting because the run completed in milliseconds; the parse failure happened on the legitimate output.

Fix

packages/server-build/src/lib/parsers.ts:

• TURBO_TASK_RE now accepts both : and # between package and task, and treats the trailing (duration) as optional.
• New TURBO_ERROR_LINE_RE / TURBO_FAILED_LIST_RE patterns parse turbo 2.x's failure summary lines ( ERROR pkg#task: ... exited (N) and Failed: pkg#task).
• Small upsert helper so a later failure line for the same task overrides an earlier cache bypass "pass" classification.
• When the Tasks: N successful, M total summary line is present, it is trusted as ground truth for totalTasks / passed / failed (failed = M − N) — preserving the passed + failed === totalTasks invariant even if per-task lines drift in a future turbo release.

Before / after

Manual repro against the captured turbo 2.9.6 output for --filter @paretools/shared --force:

Before:

{ "success": true, "duration": 1.3, "tasks": [], "totalTasks": 1, "passed": 0, "failed": 0, "cached": 0 }

After:

{
  "success": true,
  "duration": 1.3,
  "tasks": [
    { "package": "@paretools/shared", "task": "build", "status": "pass", "cache": "miss" }
  ],
  "totalTasks": 1, "passed": 1, "failed": 0, "cached": 0
}

Failure case (deliberate broken TS file under --force):

{
  "success": false, "duration": 1.4,
  "tasks": [ { "package": "@paretools/shared", "task": "build", "status": "fail", "cache": "miss" } ],
  "totalTasks": 1, "passed": 0, "failed": 1, "cached": 0
}

Test plan

• 12 existing turbo parser/formatter tests still pass (no regressions on legacy # fixtures)
• 5 new regression tests for modern turbo 2.x output:
  • cache bypass on --force
  • cache hit (replaying logs)
  • cache miss (no inline duration)
  • failing build via ERROR + Failed: lines
  • invariant check across all 10 fixtures: passed + failed === totalTasks
• Full @paretools/build suite green (262/262)
• pnpm format:check clean on touched files
• ESLint clean on touched files
• Manual repro against real turbo 2.9.6 output via the rebuilt parser confirms correct accounting

Changeset

patch to @paretools/build.