api_dispatch silently converts a text-block-free API response into an empty-string SUCCESS (7/10 Fable calls lost in #2202)

The bug

src/explore_persona_space/llm/api_dispatch.py:820 (and the identical line at :1163):

text = next((b.text for b in msg.content if b.type == "text"), "")

When an API response carries no text content block — the model emitted only a thinking block, the output budget was exhausted before any text was produced, the response was a refusal with no text, or the content list is empty — this defaults to the empty string and the item is recorded as a success:

{"result": "", "error": false, "reason": null, "category": "ok"}

No exception, no error flag, no reason, category: "ok". The caller cannot distinguish "the model returned nothing" from "the model returned an empty answer on purpose". This is a silent default swallowing a fault — the exact shape CLAUDE.md's fail-fast rule forbids ("no value placeholders, silent defaults, or fallbacks that swallow the fault; the crash IS the signal").

This is a shared-library path: every caller of dispatch_calls inherits it, not just the one that found it.

How it was found (#2202)

In #2202, 7 of 10 Fable-5 digest chunks returned empty. The failure digest was 5/5 empty; the 500-sample digest was 2 complete, 1 truncated mid-word, 2 empty. Every one of the seven was classified category: "ok". Downstream, parse_modes("") returned the empty list, so the empties contributed zero proposals and the pipeline continued as though those chunks had merely found nothing — mode discovery silently ran on 30% of its intended input and the run reached awaiting_promotion with the gap visible only as a line in the coverage-notes takeaway.

Aggravating factor at the call site: the caller passed parse_response=lambda t: t rather than parse_response_meta, so stop_reason was never persisted and the PROXIMATE cause is now unrecoverable from the artifacts. That is a caller-side choice, but the library makes the safe choice easy to skip: the metadata-aware parser is opt-in while the lossy default is free.

Why the existing gates did not catch it

• The llm-judging rule-26 pilot gate keys on stop_reason == "max_tokens", which requires parse_response_meta. With the plain parser there is no field to gate on.
• No gate anywhere treats an empty model reply as a failure condition.
• The #2202 pilot probe was a 16-token "Reply with the single word OK." round-trip, which exercises auth and routing but not the production prompt shape (the sibling-axis version of #2152 — a pilot must exercise the instrument it gates).

Proposed fix

1. Make the empty case explicit at the library boundary. When no text block is present, do NOT return the empty string as a success. Either raise, or return a typed failure record carrying category (for example empty_response), the response stop_reason, and the content-block types actually returned, so the caller sees WHY. Preserve the existing error / reason / category record contract so current consumers keep working.
2. Persist stop_reason on every path, not only when a caller opts into parse_response_meta — the field is already getattr-read at :823/:1167; record it on the result object regardless of which parser the caller passed.
3. Audit and migrate existing callers. Enumerate every dispatch_calls call site; any that can be fed a non-trivial prompt should either handle the typed empty record or opt into the metadata parser. Report the list.
4. Add a regression test that a mocked response with no text block (and one with an empty content list) produces a FAILURE record, not {"result": "", "category": "ok"} — plus a test that stop_reason survives on the default parser path.
5. Consider a lint check in workflow_lint.py for the next((... type == "text" ...), "") shape, so the pattern cannot silently reappear elsewhere.

Acceptance criteria

• A text-block-free response is a typed failure, never a success with an empty result, at both :820 and :1163.
• stop_reason is persisted on every dispatch path.
• Every existing dispatch_calls caller is enumerated and either migrated or explicitly recorded as unaffected.
• Regression tests cover the no-text-block and empty-content-list cases.
• The full relevant test suite is green on main.

Files of record

• src/explore_persona_space/llm/api_dispatch.py:820, :1163, :823, :1167
• #2202 (discovery; eval_results/issue_2202/fable_reads/, dispatcher cache data/issue_2202/fable_cache/), scripts/issue2202_labels.py:409 (the caller-side check that an empty string passes)
• #2152 (pilot gate must exercise the instrument it gates — sibling, not a duplicate: that one is about the pilot's transport, this one is about the library's silent default)
• .claude/rules/llm-judging.md rules 23/26 (max_tokens + pilot gating)