feat: add Model Lifecycle Intelligence

[pacphi]

Closes#110

Summary

• adds provenance-aware model inventory, independent lifecycle dimensions, typed compatibility/migration edges, bounded scope-keyed snapshots, conservative history diffing, and Route Intelligence invalidations
• discovers Claude, Codex, OpenCode, and Ollama through owned/configured sources without picker scraping or inference probes
• adds read-only ak models status, refresh, diff, explain, and plan; ordinary reads remain cache-only, network-silent, and token-silent
• makes Usage → Models the exclusive home for model lifecycle panels and places Models before Sessions
• presents exact owner-visible model selectors, human model providers, API rates or local/plan use, and windowed observed-use rows instead of generic private-model or access-path labels
• adds sortable route columns, a bounded human-readable consumers panel, a shared chevron pattern, counted catalogue facets, and internally scrollable tables
• replaces opaque change-history identity hashes and enum labels with a bounded semantic table naming the exact model, provider, host, plain-language change, evidence status, and detection time
• removes capabilities and lifecycle columns where they cannot establish useful facts, removes the live source-coverage panel, and documents source semantics instead
• adds direct Ollama loopback API discovery for exact installed names, safe metadata, loaded state, memory/VRAM/context, and local-compute pricing; the CLI list remains a bounded partial fallback
• removes uncited automatic retirement claims; route attention now requires direct first-party lifecycle evidence and names the route, current model, replacement, source, and exact plan command

Evidence and safety

• public catalogue discovery never becomes entitlement, policy, successful-use, or routability evidence
• observed use is windowed independently from configured routes, so GPT-5.6 Sol and historical OpenCode usage appear without manufacturing recent Ollama use
• partial, stale, failed, malformed, schema-invalid, or cross-scope discovery cannot manufacture removals
• dashboard envelopes use explicit allowlists; full Ollama templates, modelfiles, parameters, licenses, and private digests are not exposed
• Ollama HTTP discovery is loopback-only, byte-bounded, concurrency-bounded, and governed by one end-to-end refresh deadline
• kit.json and ak host pick remain the only routing-policy mutation surface
• no runtime or development dependency was introduced

Official 2026-08-25 OpenAI model pages back the displayed public API rates: GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna. No unsupported retirement notice is inferred from catalogue state.

Verification

• pnpm run check — pass
• native coverage — 86.97% lines, 79.94% branches, 84.58% functions
• pnpm run test:ui — 329/329 deterministic Playwright checks; proves Models-only panels are absent from other Usage views and the change ledger has real internal overflow
• Agentic QE integration execution — 14/14 pass
• Agentic QE SAST — zero vulnerabilities
• pnpm audit --prod — no known vulnerabilities
• ak models refresh --all --json — 440 models, 28 configured, 14 observed; all five sources complete; 15 exact Ollama models recorded under host/provider ollama
• Markdown/internal links, stable command surfaces, npm pack validation, and diff hygiene — pass

Agentic QE generic aggregation is recorded transparently in ADR-0032: its quality gate rejected retained cross-process evidence as stale, and aqe prove could not ingest the native repository run, so no artificial 98 score is claimed.

[pacphi]

Implemented the operator-first Models pass in 1791a8b: configured routes now lead; the catalog explorer is progressive; private identities use semantic labels rather than visible keyed references; source coverage and progressive details are clearer. ADR-0032 is reconciled to Accepted because the remaining source-enrichment work is not release-proved.

Validation passed: pnpm run test:ui (306 checks), pnpm run check, focused Dashboard and read-model tests.

[pacphi]

Follow-up operator pass in 294f2f2: the Models dashboard now shows exact configured or observed model selectors, names, and recorded providers in the owner-only loopback view; derives published API list rates where available; shows host plan utilization alongside the rate; removes the low-value lifecycle/capability route columns and source-coverage panel; and renders human-readable consumers in a bounded scroll pane. Credentials, endpoints, scopes, digests, aliases, and evidence identifiers remain protected. Validation passed: pnpm run check and pnpm run test:ui (308 checks).