Fix stack 04: consistent finish_reason for structured/empty/truncated results

[ebarti]

Stacked review-fix PR 4/N (base: review-fixes/03-events).

The same vendor condition produced opposite verdicts across adapters (REVIEW.md §1.1, §1.2, §1.3). Unified:

• Unsatisfied output_schema → failed everywhere. Claude previously fell through to finish_reason="done" with parsed_output=None, reporting a schema failure as success. Shared message via _common.structured_output_unsatisfied_error.
• Empty completion → failed everywhere (no output AND no tool calls AND no structured output), via _common.empty_completion_error. Codex previously failed on any empty output even after real tool work; now only when nothing usable came back. Claude/Antigravity previously succeeded on empty; now fail.
• Antigravity extracts a finish reason and maps a non-natural stop (MAX_TOKENS→max_tokens, SAFETY/RECITATION→failed) instead of reporting truncated/blocked responses as success. (ASSUMES VENDOR BEHAVIOR: probes finish_reason/stop_reason on the response and candidates.)
• Antigravity calls structured_output() only when a schema was requested, instead of unconditionally fabricating parsed_output (or crashing if the SDK errors when unconfigured).

Tests for each behavior across all three adapters. Full suite 158 passed / 12 skipped; ruff + mypy strict clean.

---

Re-review amendment. The shared structured-output failure message now claims exactly what is checked — JSON-parseability, not schema conformance (schema validation stays with the vendor SDK and the caller).

Also: added a test pinning that a SAFETY (or any unrecognized) Antigravity stop reason surfaces as finish_reason="failed" with the reason in the error, never as silent success — the unknown-reason branch was previously untested.

---

Round 2 (re-review follow-up)

• Codex turns ending in the SDK's non-terminal inProgress status — or any future unknown status — now fail closed as finish_reason="failed" with a surfaced error, instead of reading as success with partial output. Tests added for both cases.