Fix workflow initial-write exit arbitration

[dennisonbertram]

Closes #1076

Summary

Preserves the truthful source-workflow failure when a child exits before the
parent can write the initial protocol message. Initial-write failures now enter
one bounded cleanup, stdin-close, process-wait, and outcome-resolution path
instead of returning raw EPIPE before the child is reaped.

Outcome precedence is explicit:

• deadline remains first;
• semantic protocol failure remains ahead of process/transport cleanup;
• a genuine non-zero child exit and bounded stderr beat initial-write EPIPE;
• when the parent's own cleanup SIGKILL causes Wait, the original initial-write
  error remains primary;
• standalone initial-write and stdin-close failures remain visible;
• successful results and missing-result behavior are unchanged.

Failure and strict red-green evidence

Hosted PR #1070 test-race run 30660042116 failed
TestSourceManagerRunWorkflowIncludesBoundedStderrOnProcessExit: expected
child stderr diagnostic, received write |1: broken pipe.

Two deterministic lifecycle reds then proved:

1. a child can write stderr and exit 7 before the initial start write, while the
   old implementation returns EPIPE without waiting;
2. a live child can close stdin, causing EPIPE, after which the parent's cleanup
   SIGKILL could mask the original transport error.

The final tests use FIFO/flock ordering rather than sleeps, assert the exact
error precedence, and prove both child PIDs are reaped.

Verification

Exact head: fa2f7c8ed6e576006d14ec6ec989ac2d6f1a5e91, rebased onto origin/main 34c0d8627558b13851ef81ec0d25c930c045b883.

• Focused lifecycle/resolver normal stress, -count=100 — PASS.
• Focused lifecycle/resolver race stress, -count=100 — PASS.
• Complete internal/workflow normal and race suites — PASS.
• Hosted-equivalent make test-race — PASS.
• Foreground non-PTY ./scripts/test-regression.sh with Command Line Tools and isolated writable caches — PASS: normal, full race, coverage 85.6%, minimum 80.0%, zero uncovered functions.
• Repository structure and git diff --check — PASS.
• Independent cheap-tier exact-head review — APPROVED with no findings; focused acceptance/control set passed normal and race at -count=20.
• Independent Sol exact-head review — no actionable findings.
• Post-baseline-repair full exact-head regression — PASS: normal, full race, coverage 85.6%, zero uncovered functions.
• The separate provider-key and swarm-test baseline repairs are merged through Stabilize provider API-key capture test #1053 and Stabilize swarm activation control lifecycle #1047; this branch is rebased cleanly above both.

Scope and rollout

This is a narrow source-workflow lifecycle/error-arbitration repair plus its
deterministic tests and required plan/impact/log documentation. No schema,
protocol, API, cron/callback, TUI, or macOS code changes. No migration or data
repair.

Roll back if deadline or semantic protocol errors lose precedence, a child is
not reaped, stderr becomes unbounded, successful workflows regress, or
standalone transport errors are hidden. After merge, PR #1070 will be rebased
onto the repaired main and its complete local and hosted gates rerun.

Checklist

• Structured bug with complete acceptance criteria
• Plan, impact map, indexes, and durable logs
• Deterministic red-green lifecycle regressions
• Focused normal/race stress
• Full normal/race/coverage gate
• Independent cheap and Sol review
• Hosted test-fast and test-race green on this PR

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.