fix(cost): v0.40.1 — accept Anthropic dated snapshot model IDs in PriceTable.get

[cipher813]

Summary

• PriceTable.get() now retries with the trailing -YYYYMMDD snapshot suffix stripped when exact match misses, so claude-haiku-4-5-20251001 (what the Anthropic SDK returns in Message.model) resolves to the family-alias card claude-haiku-4-5 in model_pricing.yaml.
• Existing alias lookups, exact dated cards (if explicitly added), and the unknown-alias hard-fail path are all unchanged.

Root cause

morning-signal's PM 2026-05-27 firing — first edition of the public-topics-mode soak armed earlier this evening, Haiku path — crashed at cost.py:184:

PriceCardLookupError: No price card for model 'claude-haiku-4-5-20251001' active on 2026-05-28

The AM Sonnet path works because the SDK returned the family alias claude-sonnet-4-6 directly; the Haiku response carries the dated snapshot form. Pricing YAML has only claude-haiku-4-5 as the family alias (intentional, so new snapshots don't churn the card registry), and exact-match lookup against metadata.model_name failed.

Why lookup-side, not data-side

• Data-side fix (add claude-haiku-4-5-20251001 to YAML) would need to be repeated every time Anthropic cuts a new snapshot date — same bug class would recur on the next model rev.
• Lookup-side normalization is the chokepoint: one stripped retry handles every future dated snapshot across every Anthropic family. Aligned with feedback_lift_invariants_to_chokepoint_after_second_recurrence (this is the second SDK-vs-registry-naming axis after the Anthropic server-tool prefill mismatch).
• Fail-loud invariant preserved: unknown alias still raises PriceCardLookupError per feedback_no_silent_fails.

Tests

• test_dated_suffix_falls_back_to_alias — claude-haiku-4-5-20251001 resolves to the haiku card
• test_alias_lookup_unchanged — bare alias path still works
• test_exact_dated_match_wins_over_alias_fallback — explicit dated card takes precedence (operator escape hatch preserved)
• test_unknown_alias_with_dated_suffix_still_hard_fails — claude-foo-9-9-20251001 raises with full original name in error
• test_non_dated_suffix_is_not_stripped — only the strict -\d{8}$ form is normalized

Suite 888 → 893.

Test plan

• pytest -x -q passes (893/893)
• Tag v0.40.1 after merge
• Repin morning-signal to v0.40.1
• SSM pull on i-09b539c844515d549 + manually re-fire tonight's PM edition (--force) to validate the public-topics-mode soak's first edition lands

🤖 Generated with Claude Code