feat(model-capabilities): reshape match kinds to exact+prefix (Phase 1, data only)

[wpfleger96]

Summary

Phase 1 of the ratified "manifest-sourced model capabilities — no codegen" plan: a data-only reshape of the model-capability manifest. This PR carries no interpreter code — the Rust/TS interpreters, CI wiring, and generator/artifact deletion land in Phases 2/3 as follow-up dispatches. Because there is no interpreter in-tree yet, Phase 1 verification is structural, not executable (the corpus can't run in CI until Phase 2's resolvers exist; it was validated during development against a throwaway resolver matching the ratified semantics).

Two files change, both under scripts/:

• model-capabilities.json — the reshaped manifest (single runtime artifact)
• normative-corpus.json — the cross-language contract corpus

What changed in the data

• 30 exact records materialized to complete six-axis authoritative snapshots. Sparse records (24 label-only, 6 effort-override) were expanded via one final generator run out of the 3603 lift-source branch; each record carries a _provenance field naming which rule/fallback supplied its axes at materialization time. Exact records no longer inherit at runtime.
• family_rules flattened from 6 match kinds → exact + prefix only. No match_priority, no exotic segment/token grammars. Prefix semantics are boundary-aware (a token matches at string start only when the next character is end-of-string, -, or .), longest-match wins, lexicographic rule-id tie-break. The 6 openai-gpt5-* prefix rules retain a small match_aliases list of alternate literal spellings (e.g. gpt5-pro, gpt-5-6, gpt5.6) that share the rule's payload — this stays strictly within exact+prefix semantics (an alias is just another literal prefix token), no exotic kind returns.
• _generated_by removed and the root _comment rewritten to state the manifest is the runtime artifact + the maintenance rule (any capability change to a prefix rule or provider fallback must audit the exact snapshots whose _provenance names that source).
• Doc keys normalized per the verified inventory — kept as explicit documented keys or intentionally dropped; none left to be discovered at parse time.
• Corpus updated to 121 entries / 103 executable vectors (from 95/81): 81 re-pointed to the new shape, plus the gpt-5/gpt-5-pro/gpt-5-10 collision family, uncurated internal-token cases, the gptoss/gpt-neox-20b negatives, boundary vectors per distinct grammar, and two vectors guarding the highest-value pins (the class-B2 databricks_v2 multi-digit case and the sole deliberately-divergent exact record).

Curated-model immunity (the load-bearing invariant)

All 30 curated Databricks v2 records resolve at exact lookup, before any prefix matching. 30/30 resolve byte-identical (all six axes, case-insensitive path included) through the reshaped resolver, verified against the prior generated TS interpreter as oracle. Every enumerated delta below is an uncurated or adversarial input — zero curated regressions.

One curated record (databricks-gpt-5-2) carries a snapshot that deliberately diverges from what the kept gpt-5 prefix would now produce; it pins the removed dbv2-gpt-code-names-segment outcome (none..xhigh + openai-clamp-max-to-xhigh vs the prefix's minimal..high + openai-standard). The divergence is why it's pinned rather than derived, its _provenance says so, and the corpus now guards it (dbv2-gpt-5-2-divergent-pin-preserved).

Compatibility contract — match-kind flatten

Six match_kind variants collapse to two (exact + prefix). The delta is outcome-preserving for every curated model and drops only uncurated/adversarial inputs to the two provider fallbacks; each drop below names its fallback destination.

Removed match kinds → disposition (6 kinds)

Removed match_kind	Prior rule(s)	What it did	Disposition
gpt5-base	openai-gpt5-base	gpt-5 prefix guarded against a bare 1–3 digit minor version (gpt-5-10, gpt-5-2) → those fell to provider fallback	Kept as prefix on gpt-5; multi-digit guard dropped (class B1/B2)
gpt5-token	openai-gpt5-{pro,6,5,4,1}	matched the token on a left boundary with end/- right boundary → caught it embedded mid-string, treated .-suffix as non-match	Kept as prefix; alias spellings retained as match_aliases; embedded-token capture dropped (C), .-boundary widened (D)
gpt-version-segment	dbv2-gpt-code-names-segment	DBv2: routed any gpt-segment-then-numeric endpoint to openai-responses	Dropped. Curated GPT endpoints pinned exact; uncurated gpt-* DBv2 → mlflow-chat (E1/E2). Segment-anywhere isn't expressible as a prefix
segment (Claude names)	dbv2-claude-code-names-segment	DBv2: routed any endpoint containing a Claude code-name segment to anthropic-messages (omit-fields)	Partially replaced by new dbv2-claude-prefix (token claude); bare code-name segments with no leading claude dropped → mlflow-chat (F)
segment (OpenAI code names)	dbv2-sol-luna-terra-segment	DBv2: routed sol/luna/terra segment endpoints to openai-responses	Dropped. Curated databricks-gpt-5-6-{sol,luna,terra} pinned exact; bare segments → mlflow-chat (G)
segment-prefix	(unused)	declared in the enum, no rule used it	Removed, zero behavior impact

31-row literal spelling inventory (rule × alias/grammar → disposition)

Mechanical completion target = 31 literal spellings: 20 gpt5 + 6 claude/code-names + 2 gpt grammar tokens + 3 sol/luna/terra. gpt-followed-by-numeric is a grammar, represented here by its two anchor tokens.

#	Prior rule	Literal spelling	Disposition
1	openai-gpt5-pro	gpt-5-pro	retained-prefix
2	openai-gpt5-pro	gpt5-pro	retained-prefix (alias)
3	openai-gpt5-6	gpt-5.6	retained-prefix
4	openai-gpt5-6	gpt5.6	retained-prefix (alias)
5	openai-gpt5-6	gpt-5-6	retained-prefix (alias)
6	openai-gpt5-6	gpt5-6	retained-prefix (alias)
7	openai-gpt5-5	gpt-5.5	retained-prefix
8	openai-gpt5-5	gpt5.5	retained-prefix (alias)
9	openai-gpt5-5	gpt-5-5	retained-prefix (alias)
10	openai-gpt5-5	gpt5-5	retained-prefix (alias)
11	openai-gpt5-4	gpt-5.4	retained-prefix
12	openai-gpt5-4	gpt5.4	retained-prefix (alias)
13	openai-gpt5-4	gpt-5-4	retained-prefix (alias)
14	openai-gpt5-4	gpt5-4	retained-prefix (alias)
15	openai-gpt5-1	gpt-5.1	retained-prefix
16	openai-gpt5-1	gpt5.1	retained-prefix (alias)
17	openai-gpt5-1	gpt-5-1	retained-prefix (alias)
18	openai-gpt5-1	gpt5-1	retained-prefix (alias)
19	openai-gpt5-base	gpt-5	retained-prefix (multi-digit guard dropped → B1/B2)
20	openai-gpt5-base	gpt5	retained-prefix (alias)
21	dbv2-claude-code-names-segment	claude	retained-prefix (new dbv2-claude-prefix)
22	dbv2-claude-code-names-segment	opus	dropped → mlflow-chat (F); curated pinned exact
23	dbv2-claude-code-names-segment	sonnet	dropped → mlflow-chat (F); curated pinned exact
24	dbv2-claude-code-names-segment	haiku	dropped → mlflow-chat (F); curated pinned exact
25	dbv2-claude-code-names-segment	mythos	dropped → mlflow-chat (F); curated pinned exact
26	dbv2-claude-code-names-segment	fable	dropped → mlflow-chat (F); curated pinned exact
27	dbv2-gpt-code-names-segment	gpt (segment-then-numeric grammar)	dropped → mlflow-chat (E1/E2); curated pinned exact
28	dbv2-gpt-code-names-segment	gpt5 (dashless grammar)	dropped → mlflow-chat (E2); curated pinned exact
29	dbv2-sol-luna-terra-segment	sol	dropped → mlflow-chat (G); curated pinned exact
30	dbv2-sol-luna-terra-segment	luna	dropped → mlflow-chat (G); curated pinned exact
31	dbv2-sol-luna-terra-segment	terra	dropped → mlflow-chat (G); curated pinned exact

Intentional deltas (8 mechanism classes, all uncurated/adversarial)

The empirical delta measurement surfaced 8 distinct mechanism classes, not the plan's estimated 3. All are documented here rather than collapsed. Class H is an intentional correctness improvement (boundary-aware prefix narrowing an over-broad startsWith), not merely a flatten side effect.

Class	Trigger	Prior → Reshaped	Change
B1	openai gpt-5-<1–3 digit> (gpt-5-10)	fallback → base GPT-5 prefix	base guard lost
B2	databricks_v2 gpt-5-<1–3 digit> (databricks-gpt-5-10)	segment → base GPT-5 prefix; route stays openai-responses	effort/normalization narrows; route unchanged
C	token embedded mid-string (gpt-4-gpt-5-pro)	openai-gpt5-pro → fallback	embedded-token capture lost
D	.-suffixed token (gpt-5.6.x)	mixed → prefix binds on . boundary	boundary widened
E1	databricks_v2 gpt-<non-5> (gpt-6, gpt-4o)	openai-responses → mlflow-chat	segment routing lost
E2	databricks_v2 dashless gpt5-segment (databricks-gpt5-custom)	openai-responses → mlflow-chat	segment routing lost
F	databricks_v2 bare Claude code-name segment (goose-opus-5)	anthropic-messages → mlflow-chat	segment routing lost (soft degradation — loses the anthropic-messages cache opt only)
G	databricks_v2 bare sol/luna/terra segment	openai-responses → mlflow-chat	segment routing lost
H	kept-prefix token followed by a digit/letter with no separator (claude-35)	startsWith bound → fallback	boundary-aware prefix narrows an over-broad startsWith

Only E1/E2, F, G change the databricks_v2 wire route, and every one moves to mlflow-chat (Databricks's universal OpenAI-compatible wire). The corpus pins all eight classes with prior→reshaped values, including the class-B2 databricks_v2 multi-digit vector added in this PR.

Verification (structural — no interpreter in-tree yet)

• Both JSON files parse; _generated_by absent; no match_priority; match kinds are exact/prefix only.
• 30 exact records, each complete six-axis + _provenance.
• 30/30 curated records byte-identical through the reshaped resolver (oracle diff); 103/103 corpus vectors green against a throwaway resolver matching the ratified semantics (exact → boundary-aware longest-prefix incl. aliases → provider fallback, with provider canonicalization inside per Contract 5).
• Executable verification of the committed corpus is deferred to Phase 2, when the Rust + TS interpreters exist.

Originating discussion: buzz thread 30388ec72996901d324edd690681be71f49ac73f014e40776c3b1f50087fdb0a (channel tyler-wes-will).