fix: bill a failed stage for the LLM spend it already incurred

[alphacrack]

Closes #103.

stage_fail recorded no cost, so a stage that paid for LLM calls and then raised contributed $0.00 to total_cost_usd — the runs where you most want a spend figure reported zero.

The proven loss path is the distiller's grounding retry. run_distiller accumulates total_cost across up to two complete_json calls and then raises DistillError if the output is still ungrounded (distill.py) — _stage_distill never reaches stage_complete, so both paid calls vanished from the manifest. That's the failure mode this project is built around, billed at zero.

Fix, in three small pieces:

• Manifest.stage_fail(name, error, cost_usd=0.0, **meta) accumulates and recomputes the total exactly like stage_complete; the 0.0 default keeps costless failures byte-identical.
• DistillError carries cost_usd, so the spend survives the raise.
• The orchestrator's two failure handlers thread getattr(e, "cost_usd", 0.0) — any future exception can opt in without touching run() again.

Scope correction worth recording: the issue named the agent stage as the victim, but that path was already correct — _stage_normalize calls stage_complete(cost_usd=cost) before raising on blocked/failed/budget, so stage_fail only flips the status. Ingest raises only before paying. Tutorial's spend genuinely isn't recoverable (it dies inside complete_json, which returns cost only on success) and is deliberately untouched — that would need llm.py to expose partial cost, which is a different change.

Verified: 397 passed, ruff clean, mypy clean, coverage 87.53%. Revert-detection: reverting src/ fails exactly 3 of the new/extended tests, so they pin the behavior rather than describing it.

The grounding invariant — no change to what reaches published artifacts. distill.py's only edit is the exception's payload; the grounding validator, its retry, and the raise condition are untouched.

Note: #204 also touches manifest.py (adds a stage_duration helper) — different function, no overlap, but whichever lands second may want a glance.

🤖 Generated with Claude Code