feat(desktop): add pricing override management UI

[liuxiaocs7]

Summary

Makes the Usage → Pricing tab an editable, Runtime Host–backed price editor, on the production adapter path (loadPricingSnapshot / applyPricingMutation, full CAS + reconciliation). This is the M4 UI slice of #2015; #4163 is a duplicate of it.

Fixes#4163

⚠️ Product-direction note (please read)

After #2329, the maintainer re-evaluated this feature on #2218 and paused #2015 pending the #2330 override-schema contract, closing the dedicated full-catalog PR (#2218) and suggesting the surface be narrowed to overrides-only + a catalog-backed picker (the built-in set is now ~1,400 models.dev entries, so a full built-in ∪ overrides table is likely the wrong product). This PR still renders the full effective table. I've kept it current and addressed both review rounds, but I defer to the maintainers on whether to (a) narrow this to overrides-only per #2330, (b) hold it until that contract lands, or (c) close it. Flagging so the direction call is explicit rather than buried in the diff.

What this implements

• IPC: usage:pricing:load + usage:pricing:mutate. The renderer round-trips the exact snapshot it loaded as the CAS base (carries the viewed revision — no reload-latest). mutate runs through the reconciled-control path (like goal:arm): on an unknown outcome it waits for a replacement Host and reconciles against it via a non-replaying reconcilePricingMutation. The three legacy usage:pricing:{list,put,reset} handlers are removed (a test fences them — Acceptance Add RuntimeEvent read model projection #12).
• Boundary: DesktopPricing* types in src/shared with a decoder; settings.pricing.{load,mutate} scoped to the Settings-selected Host.
• Renderer: controller / pure view-model / presentational panel (mirrors the Memory page), bound to the selected Host via context. Full built-in ∪ overrides table with a Source column; resetEffect-aware reset-vs-delete copy; "Not set" vs explicit $0 cache rates; range/summary hidden on the tab; field-level validation; skeleton loading (zero layout shift); write-blocked controls carry their reason via tooltip.
• Removes the dead UsageStats.pricing read model.

Review follow-up

Round 1 (both P1s): wrong-Host write → host-scoped bridge; CAS bypass → renderer's loaded snapshot is the base.

Round 2:

• Reconnection (P1): reconciled-control path waits for the replacement Host and reconciles without replaying.
• Host-generation race (P1): mutations capture a generation epoch and drop their result if the Host re-generationed in flight.
• Conflict shows fresh authority (P1): review_required adopts the latest snapshot into the list and shows the latest value beside the draft; the notice distinguishes a real external change from an uncertain outcome (reason no longer discarded).
• Reset conflict keeps intent (P1): the confirm dialog stays open for an explicit second confirm against the latest revision.
• Uncertain/refresh-failed (P2): dims the possibly-stale list and blocks writes until reload.
• Delete copy (P2): no longer conflates "unpriced" with $0.
• Loading (P2) skeleton geometry; disabled reason (P2) via tooltip; stories (P2): dropped the FIDELITY-violating narrow duplicate and the unreachable built-in-only/custom-only states — interaction states (editor/validation/conflict/uncertain) are asserted in the renderer tests per stories/FIDELITY.md.

Correction: the earlier "waiting for #2024 / M5" note is stale — #2024 and the #2420 M5 cutover are both merged.

Still open from the review: no in-app screenshot yet (Electron isn't runnable in my sandbox — happy to attach one from a local run); large-catalog scale is the product-direction question above.

Verification

Local, all green: tsc --noEmit × 4 desktop projects + @maka/core build; biome lint . / format .; knip --workspace apps/desktop; astryx:surface-inventory; node --test — 44 pricing/usage/adapter tests including the reconcile-on-unknown IPC path, the conflict second-confirm, the uncertain write-block, and the base/host pass-through. CI (test, package, windows_recovery) green on the merge with latest main.

AI use

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: Claude Code (Claude Opus 4.8) — investigated #2015/#2218/#2330, implemented the adapter-backed rewrite and both review rounds, and the tests. Reviewed and submitted by the human contributor of record; commits carry Generated-by: Claude Code.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No

[liuxiaocs7]

Converting to draft and parking this, pending a product-direction decision.

This PR implements the Pricing editor as the full built-in ∪ overrides table. After #2329, the maintainer re-evaluated the feature on #2218 and paused #2015, closing the dedicated full-catalog PR and suggesting a narrower overrides-only + catalog-picker surface (the built-in set is now ~1,400 models.dev entries), with the override schema to be settled in #2330. So the surface this PR builds is exactly the one under reconsideration — no amount of incremental polish resolves that, and it's a maintainer call, not a code one.

The Runtime Host adapter (#2148) and Host projection are already merged; the remaining review items that are code-fixable have been addressed across the rounds (host-scoped writes, CAS against the loaded revision, reconciled-control reconnection that never replays, generation-race invalidation, conflict/uncertain UX, field-level validation, skeleton loading, disabled-reason tooltips, Heading role, reason propagation, tests). What remains open is genuinely gated on the direction: whether to keep a full effective table (needs pagination/virtualization + realistic fixtures) or pivot to overrides-only per #2330.

Happy to reshape this into the overrides-only + picker surface once the #2330 contract is agreed, or to close this in favor of that. Parking rather than closing so the adapter-backed wiring and tests are easy to pick up.