fix(headless): accept clean exit + empty parent stream for context:fork skills (#160)

[dolho]

Summary

Closes the failure mode in #160: scheduled headless tasks that invoke
skills with `context: fork` in their frontmatter were 500'ing with
"Task returned empty response". The issue documented 8 consecutive
daily failures on one agent.

Root cause

Claude Code's `context: fork` mechanism runs the skill's work in a
sub-context whose output never reaches the parent process's stdout.
The parent exits cleanly with `return_code == 0` and a populated
`result` line — so `_classify_empty_result` returns None and the
flow falls through to the end of `_finalize_headless_result`, where
the unconditional `if not response_text: raise 500` fires.

Fix

When the parent process reports clean completion (`return_code == 0`
AND `metadata.cost_usd is not None`), synthesize a placeholder
response instead of raising 500. Other failure modes — non-zero exits,
auth failures, the #520 lost-result-line case — are caught by branches
earlier in `_finalize_headless_result` and never reach this code.

Why not Option C ("check tool_count > 0")

`metadata.tool_count` reflects only parent-stream tools. Fork-context
tools don't emit to the parent stream, so they aren't counted. Using
`tool_count > 0` would defeat the purpose for the exact case the
issue reports. The clean-exit + cost-set guard is the strongest signal
we have that claude itself reported success.

Test plan

• `tests/unit/test_headless_context_fork_empty_response.py` — 5 cases:
  • Clean exit + empty parts → placeholder
  • Clean exit + real output → unchanged
  • Non-zero exit + empty parts → still errors (does NOT placeholder)
  • Clean exit but missing cost (lost result line) → 502 via classifier
  • Source-level pin that the branch + its guards survive a future refactor
• Neighboring tests (`test_empty_result_classification`, `test_claude_code_result_recovery`) still green — no contract break.
• Real-world: a fleet agent with a known `context: fork` skill runs to success via the scheduler.

Relates to #160.

🤖 Generated with Claude Code

[vybe]

Validated via /validate-pr. Two-condition guard (return_code == 0 AND cost_usd is not None) is the right signal — correctly rules out the #520 lost-result-line case. 5 test cases cover all the relevant branches including the regression pin. LGTM.