fix(extraction+gliner2): capture noise — 5 spine defects, a GLiNER2 slot collision, and Pitfall 11's broken label ceiling

[tkalevra]

A live roster turn stored phantom entities, facts about verbs, a breed filed on the user, and a person whose gender was a number. The noise had two independent sources in two different layers, and one of them turns out to be a Pitfall 11 violation.

All defects reproduce deterministically with the production spine flags on (SENTENCE_PIPELINE / SPINE_NAMING_CHAIN), and every one of them had really been written to a live memory.

Layer 1 — the spine deriver (5 defects)

#	Input	Stored (wrong)	Now
1	...named Rowan(goes by Ro)	name = rowan(goes	rowan + alias ro
2	My son Rowan (goes by Ro) is 12.	("goes", age, "12"), alias lost	(rowan, age, "12") + alias
3	Ines is my wife.	(wife, spouse, user) — phantom person	(ines, spouse, user)
4	Quinn, age 10, F.	(f., instance_of, age)	(quinn, age, "10")
5	My dog Bracken is a morkie.	(user, "dog", "morkie"), breed unbound	(bracken, instance_of, morkie)

1. A bracket is a spaCy prefix/suffix rule, so it never fires mid-token — the glued run survived as one PROPN and became the name. _install_bracket_infix adds the missing infix rule; spaced and unspaced forms now parse identically.
2. A parenthetical has no subject, so spaCy re-attaches the alias verb as nsubj of the following copula. _binding_nickname read only the relative-clause form; it now also accepts the linearly anchored parenthetical (vocabulary from the growable alias_predicate cue map), the measure chain rejects a non-nominal subject and recovers the real one left of the bracket.
3. _bound_name_for_type read only my <role> is <Name>, never the reverse. Binding branch (4), gated on 1st-person possession, casing-robust so capture is not name-dependent.
4. A type noun already saturated by its own cardinal nummod cannot host a name; a single-character name is an initial; the appositive chain stores the pair as the scalar it is.
5. Binding branch (5) binds the determiner-introduced predicate nominal to the subject NP's name; the attribute-scalar chain steps aside for exactly that shape — narrowed by the complement, so a real possessed scalar (my address is 123 Main Street, …) is still captured.

Also excludes the naming noun name from the type↔name binding and the copula-name chain: my mothers name is Priya belongs to _chain_genitive_name, and reading it as a type stored (priya, instance_of, "mothers name").

A regression this PR caught in itself: splitting the bracket promoted the nickname to a real token, so it became the nearest preceding PROPN and he resolved to ro — the age landed on the alias. Found by running the real /ingest pipeline, not by the unit tests. An alias is a second name, not a new referent, so the coref scan now skips the alias run (already marked in _ni_suppress).

Layer 2 — GLiNER2 (found only by running the live pipeline)

Slot collision. GLiNER2 scores each candidate rel_type independently, so nothing stopped it returning the same (subject, object) under two labels. Verbatim raw output for "My son Rowan is 12.":

{'age': [('Rowan','12')], 'has_gender': [('Rowan','12')], 'height': [], ...}

Both committed → has_gender = 12, a person whose gender is a number. Same shape as the gender named F / quantity is 10 F / 12 m rows in the live store. Resolved with metadata the schema already carries (migration 101's scalar_datatype): a bare cardinal satisfies integer, is not a quantity (needs a unit), and is only vacuously a string — so age wins and has_gender is dropped. Non-numeric spans are left alone, because pref_name + also_known_as legitimately co-claim a name.

Pitfall 11's other half — the per-instance label ceiling. The label-content rule was honored. The count rule was silently lost when the caps became per-pool. GLiREL states it outright:

"We limit the number of relation type labels prepended to each training instance to 25." (arxiv:2501.03172)

The comment directly above the cap constants says exactly that, and _GLINER2_RELATION_LABEL_CAP = 25 still sits there marked "legacy". But scalar 15 + relational 20 are merged into one dict for a single extract_relations call — so the model received 35, 40% beyond the training maximum. Measured live: 33 labels, and that is the call that fired the colliding labels. Only /ingest breaches it; the other three call sites use the single capped helper (≤20).

The paper also measures the cost of merely approaching the cap — F1 by candidate count m: m=5 → 94.20, m=10 → 87.60, m=15 → 84.48 (FewRel). Fewer labels is measurably more accurate, not merely safer.

_enforce_gliner2_total_label_budget trims at the one point both pools are still distinguishable. Relational is sacrificed first (a scalar rel is the more specific storage path) but never below a reserved floor, because "a large scalar pool crowds relational out entirely" is the exact failure the split budgets were introduced to fix. Truncation is logged with the dropped label names — a silent cap reads as "we considered everything" when we did not.

Verified live after the fix: candidate_set_built total=25 (was 33), and a clean-tenant ingest of "My son Rowan is 12." now stores exactly one attribute row: rowan | age | 12.

Not changed here: the relational pool's own cap of 20 is within the training distribution but sits in the measured degradation zone. Pitfall 11 says a label-set change is a benchmark-gated whole-label replacement, so lowering it needs a benchmark, not a silent edit.

Design

Subject-agnostic throughout — dependency shape, morphology, orthography, and growable DB metadata/cue classes only. No name, role, breed, gender, or rel_type literal anywhere in the change.

Tests

388 passed. New: 14 spine regression tests, 9 GLiNER2 collision tests (driven by the verbatim payload captured off the line), 13 label-budget tests — plus the 3 tests in test_spine_genitive_name_and_alias.py that were already red on main.

Regression sentinels included deliberately: the possessed attribute scalar (an earlier, broader guard silently dropped it), an unpossessed predicate nominal, a real role apposition, non-numeric co-claiming, and identifier tokenization (CVE-2024-9999, x86-64, 10.0.0.5:8080, AB123-XY all verified whole; contractions unaffected).

🤖 Generated with Claude Code

https://claude.ai/code/session_017uPqpFMSnoi1F4EXzKcPSW