fix(inspector): expose usage and context evidence

[phodal]

Summary

• Add one provider-aware Usage and context disclosure to standalone Harness Inspector and Studio Session Detail.
• Normalize Codex cumulative token snapshots, preserve Claude cache-creation tokens, and attach matching Cursor Context Usage Canvas evidence.
• Keep Codex developer/system/reasoning text and raw context items out of dialogue and portable reports.

Why

• Issue/Story: None; justified maintenance from a live Harness Inspector evidence audit.
• User or maintainer outcome: Inspector can answer which token and bounded context evidence was observed for Codex, Claude, and Cursor without inventing billing totals or exporting prompt text.

Traceability and Scope

• Spec/ADR, if applicable: docs/specs/2026-08-28-harness-inspector-usage-context.md
• Acceptance criteria addressed: AC-1 through AC-7.
• Canonical owners changed: Codex, Claude, and Cursor session adapters; shared session summary/report projection; standalone Inspector and Studio Session Detail.
• Explicit non-goals: raw prompt/context export, billing or account claims, a stable public host-schema contract, or a universal provider transcript IR.

Change Type

• Feature
• Bug fix
• Tests only
• Documentation/community
• Refactor with no intended behavior change
• Dependency, packaging, or infrastructure

Test and Review Evidence

Check	Result
npm run check on Node 24.19.0	Passed: root 1548 passed, 2 skipped; Harness 172; Harness UI 31; Studio 293; generated sources and package/archive verification passed
npx vitest run test/sessions/session-analysis-providers.test.mjs test/sessions/session-analysis.test.mjs test/sessions/commit-session-link.test.mjs	Passed: 143 tests
npx vitest run test/reporting/harness-inspector.test.mjs	Passed: 29 tests
npm run test:browser -w @qoder-ai/harness-studio -- --grep "opens a project workspace and compares Inspector-discovered Sessions"	Passed: 1 Playwright scenario
node scripts/harness-inspector/visual-contract-check.mjs	Passed: 15 surface and viewport states

Manual or visual evidence:

• Studio usage/context detail checked at 1440x900, 1024x768, and 390x844 with no console/page errors or document-level horizontal overflow.
• Standalone Inspector checked across wide, compact, and narrow layouts with no clipped facts, below-floor targets, page errors, or horizontal overflow.
• A live local render found 58 sessions and 7,989 tool calls; usage was observed for 44 of 45 Codex sessions and all 13 Claude sessions. This workspace had no matching Cursor Session, so Cursor context matching is fixture-backed.

Risk and Recovery

• Compatibility and cross-platform impact: filesystem work uses node:path and temporary-directory fixtures; Windows and Linux remain authoritative through PR CI.
• Package, plugin, schema, or generated-file impact: additive report/session fields only; no generated files or dependency changes.
• Rollback or recovery path: revert commit 13d5f68.
• Residual risk or unverified boundary: local provider files are private implementation formats and may drift; Cursor live-session matching was not available in the current workspace.

AI Involvement

• Level: Generated by Codex, with explicit Co-authored-by metadata.
• Human review and validation: Human review is pending in this PR; automated, browser, visual, and live-host evidence is recorded above.

Checklist

• I followed AGENTS.md, CONTRIBUTING.md, and the relevant canonical-owner guidance.
• The change is focused and does not include unrelated local or generated state.
• Tests and documentation match the behavior actually delivered.
• Markdown links were checked when documentation moved or changed.
• Cross-platform behavior was considered for Windows, macOS, and Linux.
• Package/runtime verification was run when shipped files or dependencies changed.
• User-facing or compatibility changes are recorded in CHANGELOG.md. Intentionally unchanged because repository instructions prohibit proactive release-metadata edits without explicit authorization.
• I have the right to contribute this work under the repository MIT License. Maintainer confirmation requested.

[phodal]

Follow-up: per-response usage and context-window evidence

Each retained model inference now appears in Turn order as Model response N:

• provider-observed token breakdown for that response
• same-event context used / window / percent full when available
• explicit not observed for this response when the provider did not emit a window

Live local evidence: Codex projected 5,341 rows across 44/45 sessions, all with same-event context-window evidence. Claude projected 981 rows across 13/13 sessions, with no context-window field available. Cursor historical rows remain evidence-gated; the current Context Usage Canvas is not copied onto older responses.