What happened

Two of twelve identity-bench rounds ended with every document done, 30+ entities, 50+ facts, and every entities.type_id NULL, so all 21 anchors (located by class) were missing. The server log had no WARN.

The surviving KB (guard-run7 reverse) shows 13 align_types jobs, all done, no last_error, each taking 10–50 s, and zero type_bindings rows. Reproducing with the same entities against SiliconFlow deepseek-ai/DeepSeek-V3.2 showed why: at temperature 0 the model often answers the kind-word prompt not as the documented {"b": [[id, key]]} but as an id-keyed object, in several variants:

{ "0": null, "1": "organization", ... }
{ "0": ["organization"], "4": [null] }
{ "0": ["organization", "机构 — 研究院、研究所、集团…"] }
{ "0": [0, "organization"] }

parse_kind_word_response read any object without a b array as "no pairs, nothing malformed", so every item counted as unanswered. One unanswered vote out of two means no decision, nothing is written, nothing is logged, and nothing is re-enqueued (only call/parse failures were). The same prompt returns the same shape every round, so the base stayed untyped through all twelve documents. The "good" sibling round only bound anything on 2 of its 12 runs, for the same reason.

What changed

• utopia-extract/align.rs: the parser reads all the shapes above, plus b as an object, {"id":..,"key":..} pairs and a bare top-level array. Unreadable ids/keys count as malformed instead of absent.
• type_alignment.rs: a pass that parses to zero choices is treated as a failed pass and logged at WARN with finish_reason, char count and a snippet of the reply. The completion line reports unanswered, and a WARN follows when it is non-zero. Unfinished runs (failed or unanswered) re-enqueue themselves with a reask counter in the payload, at most MAX_REASK (3) times with a growing delay; new words or stale bindings still re-enqueue as before.
• main.rs: the job dispatcher passes reask through.
• Tests: four parser tests for the shapes; an_id_keyed_reply_still_binds_the_kind_word reproduces the all-NULL case through the scripted model and asserts the class lands; an_unreadable_reply_is_asked_again_a_bounded_number_of_times asserts no decision, one reask=1 job, and no more once the budget is spent.

Resolution, adjudication and governance are untouched.

Verification

• cargo test --workspace, cargo clippy --workspace --all-targets -D warnings, cargo fmt --check green locally against a dedicated database.
• Copying the broken KB's 29 entities into a fresh base: old binary binds 0 words, this branch binds 18 (typed 21/29).
• Full identity bench, one round each order, SiliconFlow DeepSeek-V3.2 + bge-m3:

Earlier rounds had 3–9 anchors missing. The f1 numbers are not directly comparable with those rounds: with more anchors usable, the pre-existing INST/LN/P2 splits now enter the score (pair-for-pair identical to guard-run7 forward, which also had 2 missing). Those splits are candidate-generation gaps on the identity side, out of scope here.

Follow-ups noted separately: the phrase-binding parser has the same silent path; scripts/e2e_type_drift.mjs is stale (documents.find on a {docs: [...]} response) independent of this change.

🤖 Generated with Claude Code