fix(registry): bedrock records for the current Claude generation (#607)

[gadievron]

What

Adds five Bedrock records to config/models.json for the current Claude generation, closing the gap where a config naming one of these models under provider bedrock priced real tokens at $0 with cost_incomplete:

id	input	output
anthropic.claude-sonnet-5	2.0	10.0
anthropic.claude-opus-5	5.0	25.0
anthropic.claude-fable-5	10.0	50.0
anthropic.claude-fable-5-1	10.0	50.0
anthropic.claude-opus-4-7	5.0	25.0

The last two come from a sibling sweep over the same gap class (current Claude models served on Bedrock with a direct record but no bedrock record) — fixed here rather than left as a known gap.

Why these ids (and not us./global. twins)

The issue's original direction — mirror the 4.x-era us./global. inference-profile records — is superseded by the vendor's current documentation:

• Anthropic's current Claude-in-Amazon-Bedrock integration serves the -5 generation (with Opus 4.7/4.8) under plain anthropic.claude-* model IDs (Supported models table, fetched 2026-09-21).
• The legacy guide omits these models from its inference-profile table explicitly "because they do not have ARN-versioned model IDs".

So there are no us./global. ids to mirror; adding them would fabricate ids. The regression test pins this: it fails if a us./global. twin for this generation is ever added.

Pricing basis

• Prices equal the direct Anthropic rates, re-verified 2026-09-21 against Anthropic's pricing page (Sonnet 5 $2/$10 — the page's note confirms the launch price is now the standard price; Opus 5 $5/$25; Fable 5 and Fable 5.1 $10/$50; Opus 4.7 $5/$25).
• Bedrock global-endpoint pricing carries no premium; the 10% regional premium is an endpoint choice, not a model-id variant, so id-level pricing carries the global rate (same modeling as the direct-provider records, which likewise don't distinguish inference geography).
• Each record's source carries its receipts. Honest limitation, stated on every record: no live-account list-inference-profiles verification was possible from the authoring environment (no AWS credentials) — the ids are vendor-documented, not account-verified. The issue's evidence-state comment asked for exactly that check if reachable; it was not, and the receipts say so rather than implying otherwise.

The cross-check learns the new id shape

test_models_price_cross_check.py's family normalization now strips a plain anthropic. prefix (region prefixes were already handled), so the new records are guarded by the same-family price-agreement invariant.

Testing

• New regression test (test_issue607_bedrock_current_generation.py): resolution under bedrock at the direct rate (map-compare against the anthropic records — no rate literals duplicated), provider-scoped existence via find_model, the no-fabricated-twins guard, and the alias shape.
• RED confirmed on the base (2 tests failed with the records absent; stash-restore re-check: reverting only the JSON turns them RED again — not vacuously green).
• Full related suite green: 104 passed, 1 skipped (registry, price cross-check, A missing config/models.json record silently prices a model at zero #216 unpriced-loud, model pricing lookup is exact-match on id, so a config spelling a model differently than models.json reports $0.0000 for the whole run (claude-haiku-4-5, google/gemini-3.7-flash observed) #434 alias, config/models.json lacks direct anthropic-provider records for claude-sonnet-5 and claude-fable-5 #601 completion, LLM registry, A budget-exhaustion empty completion (finish_reason='length') is still retried as transient — the message now states the deterministic cause, but the classifier treats it as recoverable #569 budget retry, Pip floor-vs-pin drift between pyproject.toml and requirements.txt has no consistency check #428 floors, bedrock adapter, plus the new file).

Deferred (same class, on record)

• The Mythos family: no documented plain-id serving on the current integration (glasswing-gated; only anthropic.claude-mythos-preview appears, invitation-only, and it has no published price row anywhere).
• A plain anthropic.claude-opus-4-8 record: the family already carries the live-verified us./global. twins; whether the plain id is also served deserves its own verification pass rather than an assumption here.

Fixes #607

Coordination

• feat(cost): provider-aware cache accounting — the prerequisite for any prompt-caching control (#626) #661 (cache accounting): the direct-anthropic entries there gain additive cache multipliers these bedrock records do not carry — the rate-agreement test pair-compares (input, output) so both PRs merge in either order. A cached run on these bedrock records takes feat(cost): provider-aware cache accounting — the prerequisite for any prompt-caching control (#626) #661's unpriced_cache_models incomplete path (honest — AWS-side cache pricing unverified).

[gadievron]

Coordination note (the process's A4-final re-fire, 2026-09-21): this PR composes with #661 (the cache-accounting change) — the direct-anthropic entries there gain additive cache multipliers these bedrock records don't carry, so this PR's dict-equality test would break on the merged tree while the rates still agree. The latest push pair-compares the (input, output) rates instead (the same update #661 made to its own tests). Also for the record: a cached run on these bedrock records (no cache multipliers yet — AWS-side pricing unverified) takes #661's unpriced_cache_models incomplete path, which is the honest behavior until the rates are sourced.