Refresh nine model landscape checks after UTC rollover - #1124

Scope

UTC rollover on October 15 made nine October 7 model landscape checks eight days old, causing every PR's current-date validator to fail. Refresh those nine checks from current primary cards, release notes, publisher searches and consumer source; retain model pins and the 42-day/7-day policy.

• Base commit: a339425ca56ec3fb87dff0fa8ad5dffc83354168
• Lane: lane:foundation
• Owned paths: catalogs/foundation/model-currency.json, tests/test_model_currency.py, docs/decisions/2026-10-15-model-currency-landscape-check.md, and their manifests/evidence.json bindings.
• Material new finding: Anthropic now lists Haiku 5.5 as Active (latest), released October 7, API ID claude-haiku-5-5. Record the successor and triggered migration-review gate explicitly; the old 4.5 pin is not represented as current-leading or silently replaced.
• Current cards and upstream qmd/MinerU/Magika source support bounded retention/qualification gates for the other roles. Re-read candidate cards, update the alpine candidate head, and remove elapsed October 6 cooldown rationales.

SOTA sources

All primary sources were read on 2026-10-15 UTC; the nine inventory rows and decision table enumerate role-specific URLs and available exact revisions.

• NVIDIA selected model card at d1f2f257 and Google's EmbeddingGemma 2 card at 914f7f89.
• tobi/qmd 93d211f9, src/llm.ts: actual reranker/expander defaults and loader; selected and alternative publisher model cards are linked in the rows.
• MinishLab Potion card at e9d2a44c and current miru tokenlearn/dual cards.
• OpenDataLab/MinerU b7f3dd0c, registry.py: GGUF/mmproj selection, with official base and alpine candidate cards.
• Kompress e00cc5e6 card, pinned/current ONNX metadata, and ModernBERT card at 8949b909: selected artifact identity and prescribed tokenizer pairing.
• google/magika b1909890, model changelog and primary PyPI release metadata.
• Anthropic Haiku 5.5 specification, current models overview and release notes: released successor and exact API ID.
• Existing supported scripts/freshness_propose.py --model-currency on the stated base: primary metadata observations and all nine expired refresh proposals; it proposes review rather than making landscape decisions.

Evidence-class table

Local commands run

python3 scripts/freshness_propose.py --model-currency --state-dir /var/tmp/currency-model-currency-incident-proposals-20261015
rc0; all nine refresh_landscape proposals, current primary release-line observations

model_currency_problems(base_inventory); model_currency_problems(updated_inventory)
actual current UTC date, no date argument: 9 errors -> 0 errors

python3 -m unittest tests.test_model_currency tests.test_model_currency_propose
rc0; 46 tests, including real-row Haiku successor regression

python3 scripts/validate.py
rc0; 71 components, 12025 hashed files, 240 receipts, 4 profiles

git diff --check
rc0

Decision record

docs/decisions/2026-10-15-model-currency-landscape-check.md records the source-review method, each role's evidence, alternatives and gates. No model-quality, native inference or production configuration qualification is added. Haiku 5.5 migration review remains outstanding. A separate age-triggered refresh follow-up starts after this incident PR lands.

Host evidence

No evidence/hosts/ changes.

Checklist

• No workflow, permissions or billing changes.
• No credentials or private account/state inputs recorded.
• Peer worktrees and files preserved.
• Current-date policy retained without override or relaxed limits.

Follow-ups

1. Astra P3: retain primary-source captures. Retain referenced public source bytes and acquisition dates, hash-bind them, and add missing exact repository pins to the mirror inventory so the nine source judgments can be independently reproduced. No unverified source-review claim is relabeled as independently verified.
2. Astra P3: complete future reviewed refresh. Use the existing day-five proposer plus sourced human/agent review in the separately queued scheduled-refresh follow-up. October 15 checks become due October 20 and expire October 23 if unchanged; the existing timer does not perform that review.
3. Astra P3: long scratch-path notice test portability. Make the adjacent tests/test_active_model_currency.py:947 assertion account for documented 160-character notice truncation, or use a short permitted fixture root. It is a pre-existing portability issue outside this fix round.
4. Astra P3: historical pending locator. Resolve evidence/artifacts/eval-install-repairs-20260926/gpt6-cross-family-review.md:5's explicitly pending manifests/evidence.json:29554 locator against its original reviewed commit, rather than treating it as a current-tree location.

These four follow-ups are recorded without implementation in this fix round. The committed decision's regression receipt records the fail-first historical-row fixture and corrected-row pass. The same-round P2 additions below refine the source-review prose; selected model pins, dates and policy remain unchanged.

Additional CC micro follow-ups (recorded, not implemented here):

• Headroom package row: refresh the enclosing packages row from 0.40.0 to the newly released 0.41.0 through its normal package review. Both Kompress model rows now explicitly record that 0.41.0's model/tokenizer pins equal 0.40.0's.
• CC P3-3: list the newer maintainer 0.6B expansion GGUF explicitly in newer_candidates; its current prose lead remains recorded.
• CC P3-4: replace the two remaining raw/main source locators with their exact available model-card revisions.
• CC P3-5: reconcile the earlier decision's still-required verification wording with the already reported results, and link the separately queued age-triggered sweep. Next expiries are October 22 and October 23; the sweep must fire before October 22, following row check age and the existing refresh lead.

The expanded same-round P2 correction records current Haiku 4.5 legacy/retirement/notice/migration state and the CC-confirmed migration owner, full-harness api-actions paired check, tracking record and readiness trigger. It also compares actual 0.40.0/0.41.0 Headroom wheels and cites package source lines: HF_MODEL_ID:47, _PINNED_REVISIONS:73–79. Selected model pins, check dates and policy remain unchanged.