Group dashboard projects and report context limits across hosts

[pacphi]

Projects were presented as flat lists, and the Runtime context cards mixed model capacities with configuration and unavailable session information. This change groups projects from Git evidence, keeps session origin independent, and reports cached context limits using each host's actual semantics.

• System and Maintenance Projects associate evidenced worktrees with repositories while preserving existing totals, filters, identities, and navigation. Maintenance cards align the folder icon with the name, reuse the sidebar kind badge, and show all language icons. Unclassified origin remains filterable without repeated placeholder text.
• Intelligence uses matching alphabetized picker and table subgroups for repositories, worktrees, user-level learning, and other locations. Each table subgroup shows five rows before scrolling inside a bounded panel. Labels are plain names; all rows, selected keys, learning history, and KPIs are preserved.
• Usage → Score → Projects restores a plain top-10 Git-project spend ranking for the selected timeframe. Verified worktree spend rolls into the real parent project; standalone worktrees, user-level and unclassified directories stay out of the ranking. Overall totals remain intact.
• Context configuration combines the Codex rows and reuses evidenced Claude/OpenCode model inventory caches. Native versus kit-managed controls are explicit; no live-usage placeholders, repeated limitations, or polling-time CLI/API calls. The model-inventory link navigates to Usage → Models. Historical Context cards distinguish input-only, partial paired coverage, missing measurements, and no sessions; absent token/window values render as an em dash instead of zero.

Design: ADR-0050. Git common-directory/backlink evidence establishes association; shared remotes and names alone do not. Model capacity, configured requests, calculated usable windows, and session facts remain distinct.

Validation

• npm test: 3,829 passed, six skipped; 91.98% line and 80.75% branch coverage, followed by all legacy script suites passing.
• Typecheck, ESLint, complexity ceiling, Markdown lint, and build passed.
• Five focused real-Chrome suites cover grouping/counts, worktrees, origin filters, all language icons, Intelligence optgroups, empty/unknown/legacy states, keyboard navigation/focus, mobile layouts, bounded model tables, and the model-inventory link.
• Independent bounded reviews found no blockers in identity/origin, context evidence, Maintenance API privacy/counts, or Usage aggregation.
• npm run test:ui: all 491 dashboard checks and five focused browser suites (six test cases) passed. GitHub CI is green on 0acebf2: all nine OS/Node test jobs, quality, browser, internal links, and maintainer devcontainer checks passed. The consumer release devcontainer was intentionally skipped.

Screenshots

Deterministic fixtures only. Full screenshot index and reproduction.

Remaining boundaries

Claude historical transcripts lack paired windows, older Codex telemetry may be totals-only, and OpenCode may have no sessions in the selected window. Catalog limits are not used to fill these gaps.

Older footprint snapshots need remeasurement for new identity metadata. Usage cache schema 20 performs one normal reparse; original transcripts remain untouched. Desktop attribution is a recorded declaration, and histories outside discovered roots remain outside coverage. Cached limits are not live-session measurements. Only Codex currently has kit-owned context writes; Claude/OpenCode expose native controls and evidenced catalog limits.

Based on merged remediation #209 (2cae143). Existing worktrees and unrelated changes were preserved. This PR is not authorized for merge. Comprehensive documentation alignment is queued separately in #211.