refactor(#276): extract DashboardSession — phase 3b

[BorisTyshkevich]

What & why

Part of #276 — Phase 3b, completing Phase 3 with #281 (3a). PRs so far: #278 (0–1), #279 (2), #281 (3a).

DashboardSession (src/ui/dashboard/dashboard-session.ts, 36-test spec; 100/94.9/100/100) owns the dashboard execution runtime: the 6-way tile pool, wave generations reserved at wave creation (the race-closing invariant the DOM suite asserts to the exact request count), per-slot AbortControllers, runFavoriteSource's gate/stream/classify core, filter waves + curated-field merge (persistence via injected ports), runAffected's cascade, and retryFilter. Constructible without App; the input is an explicit DashboardRuntimeInput built by the shell from the favorites list — a stored DashboardDocument can replace that source later without touching the session (issue §4 requirement). Every DOM write stays in the shell behind hooks; slots keep their route's own DOM for now (session-state/shell-DOM split deferred to Phase 5 — deliberate, documented in the module header). Tile supersede remains client-abort only, never a server KILL (intentional, preserved).

Strongest signal: renderDashboard(app) keeps its signature as the integration entry and tests/unit/dashboard.test.ts passed unmodified — every streaming/race/filter-wave/recents behavior pin holds against the extracted runtime.

Teardown (issue acceptance: "tests prove cleanup"): destroy() aborts all in-flight tile/KPI/filter work, tears down chart instances, disposes the filter bar, and flips a destroyed flag making runAll/runAffected/retryFilter no-ops (re-checked after the token-preflight await) — an orphaned filter-bar debounce timer firing post-teardown can never issue a request or trigger a sign-out. buildFilterBar now returns { el, dispose() }; both the dashboard shell and the detached Data view dispose the previous bar on re-render, closing a real pre-existing orphan-timer gap. Like WorkbenchSession, destroy() has no production caller until Phase 5's route shells (the dashboard is its own tab today) — documented.

Review round (3 angles + e2e): the port-fidelity angle caught a genuine regression — with progress routed through the live slot.loadLabel, a superseded request's last buffered chunk could corrupt the newer wave's label (the old closure-captured label was immune). Fixed with a stale-generation guard before the progress hook + a unit test modelling the buffered-chunk-after-abort window. Also fixed: destroy-during-preflight window, two stale doc comments. Cross-file + conventions angles otherwise clean.

Verification: full gate green (106 files / 3,090 tests + tsc + check:arch, ui/dashboard rules active), build green, e2e chromium+webkit fully green, firefox 40/40 serialized (parallel-run timeouts are this machine's documented flake).

Acceptance criteria this completes (with #281): Workbench and Dashboard have separate route-scoped sessions and cancellation state ✅; both expose destroy() with proven cleanup ✅; route sessions do not import one another (enforced) ✅; dashboard runtime input is an explicit interface, not a favorites-list assumption ✅.

Checklist

• npm test passes (the per-file coverage gate is non-negotiable)
• Tests added/updated in the same change as the code
• npm run build succeeds (single-file dist/sql.html)
• Layers kept honest: pure logic in src/core/, network in src/net/ (injected fetch), DOM in src/ui/
• No new runtime dependency (or it's a deliberate, justified addition — see CONTRIBUTING)
• README / CHANGELOG.md ([Unreleased]) updated if behavior or the deployed surface changed
• Reconciled affected tracked work (roadmap Roadmap to 1.0.0 #68, the issue body, ADR/CHANGELOG) if this change reshaped it

🤖 Generated with Claude Code