Make the Usage report decision-first and keyboard-correct

Traceability

• Spec ID: inspector-usage-report-clarity
• Status: Implemented

Intent

Make the standalone Inspector and Harness Studio Usage report answer the reviewer's primary question first: how much of the latest observed context window was occupied, when that static snapshot was current, and how the Session reached it. Remove controls, metric substitutions, and visual encodings that obscure or misstate that decision, especially for keyboard and narrow-screen reviewers.

Acceptance Scenarios

• AC-1: The detailed Usage report visibly identifies itself and places one complete Static snapshot freshness statement beside the latest observed context value. The statement remains readable at wide, compact, and narrow widths and is not available only through clipped text.
• AC-2: Summary positions retain stable semantics across providers. Context health, input reuse, Session processing availability, and model-call evidence do not replace one another with unrelated metrics such as peak context. Provider/runtime/provenance methodology is available through one secondary, collapsed disclosure after the primary summary.
• AC-3: When all retained responses fit in the focus window, the report renders one context chart and omits the duplicate Overview, range sliders, and unavailable Previous/Next window actions. Longer Sessions retain the bounded Overview and independently adjustable Start and End controls.
• AC-4: Start and End range inputs preserve native keyboard operation without changing the selected Response. Interactive charts expose control semantics and describe the actual Arrow-key selection model; Escape clears a local selection. Returning to Trace restores keyboard focus to the Trace view.
• AC-5: Narrow layouts keep evidence values, chart labels, range controls, and response rows readable and operable. Meaningful text remains at least 12px, touch-oriented targets are at least 44px, and no document-level horizontal overflow is introduced.
• AC-6: Cache reuse explicitly states that cached input still occupies the context window. Codex context-layer item counts render as a count inventory, not as a proportional bar that resembles token-weighted composition.
• AC-7: Standalone Inspector and Studio retain equivalent terminology, information order, conditional explorer behavior, and accessibility semantics. Existing provider accounting values, progression points, prompt links, privacy omission, compaction evidence, and Context Window arithmetic remain unchanged.

Non-goals

• Changing provider adapters, usage arithmetic, Context Window inference, compaction detection, or the portable report schema.
• Adding live refresh, billing, cost, savings, or optimization verdicts.
• Completing the full standalone light/dark theme migration or adding a theme selector; this change uses the existing semantic tokens while removing the dashboard-like Usage layout.
• Reconstructing token-weighted context categories when a provider retained only layer item counts.

Plan and Tasks

1. Recompose the Usage lead in both renderers around a visible title, latest observed Context Window, full static freshness, and stable supporting facts.
2. Move runtime/accounting/provenance facts into a shared collapsed evidence disclosure and make long values wrap at narrow widths.
3. Make Overview/window controls conditional on a genuinely bounded focus subset; retain one focus chart, response detail, and dense response table for short Sessions.
4. Correct standalone key routing, chart control semantics, truthful keyboard help, and Trace focus restoration.
5. Remove item-count proportion bars, state the cache/context boundary, and tighten wide/compact/narrow styling without changing evidence values.
6. Update behavior, Studio browser, and standalone visual-contract assertions; regenerate the documentation routing graph.

Test and Review Evidence

• AC-1/AC-2/AC-6/AC-7: focused report-model and renderer assertions inspect visible copy, stable labels, evidence disclosure, cache boundary, and the absence of count-as-composition bars.
• AC-3/AC-4: Studio Playwright and standalone visual checks cover a short Session and a long Session, native range-key behavior, selection behavior, and focus restoration.
• AC-5: the Inspector visual contract must pass at 1440x900, 1024x768, and 390x844 with no clipped meaningful text, below-floor type, unreachable controls, console/page errors, or document-level overflow.
• Regression: run focused Inspector Vitest files, npm run harness-studio:test, the focused artifact-host Playwright scenario, node scripts/doc-link-graph/cli.mjs skills/better-harness, and the doc-link graph test.
• Risk: standalone and Studio own parallel renderers. Review their emitted labels, conditional branches, keyboard paths, and responsive screenshots as one change; a pass in only one host is insufficient.

Validation Results

• npm test: 107 files passed; 1,598 tests passed and 2 skipped.
• npm run harness-studio:test: 65 files and 512 tests passed after a clean Studio build.
• Focused Studio Playwright: the Inspector Session/Usage/Replay scenario passed, including the short-Session branch and Trace focus restoration.
• Standalone visual contract: capability, date, Session Trace, Usage, and Replay passed at 1440x900, 1024x768, and 390x844 with zero document overflow, below-floor type, clipped text, or console/page errors. The Usage check also exercised native range-key adjustment without changing the selected Response.
• Documentation link graph: regenerated and its 8 focused tests passed.
• Preview smoke: /health returned ok; /canvas-module.js returned HTTP 200.