feat: unify dashboard observability and provider UX

[pacphi]

Summary

This PR turns the former Live surface into a provider-consistent Observability experience, reorganizes the dashboard into three stable product areas, fixes session identity/provenance reporting, and brings the architecture and user documentation in line with the implementation.

Dashboard information architecture

• Reduce the primary navigation to three durable areas:
  • Overview
  • Usage
  • Observability
• Move Summary, Hosts & Routing, Providers, Runtime, and Intelligence into a fixed, left-aligned Overview secondary rail.
• Keep Usage secondary tabs—Scorecard, Limits, Findings, Sessions, and Transcript—in the same stable rail position.
• Present Observability Live and History as the same secondary-tab pattern used elsewhere.
• Add a heading and concise explanation to every dashboard view.
• Introduce canonical hierarchical routes:
  • #overview/summary
  • #overview/hosts
  • #overview/providers
  • #overview/runtime
  • #overview/intelligence
  • #usage/score
  • #usage/limits
  • #usage/findings
  • #usage/sessions
  • #usage/transcript
  • #observability/live
  • #observability/history
• Preserve arrow-key navigation and stable tab placement across viewport sizes.

Observability navigation and modes

• Rename the Live product area to Observability throughout UI, routes, docs, ADRs, and DDD references.
• Separate current evidence from retained evidence:
  • Live lists only projects and root sessions that are currently present.
  • History lists retained projects and sessions without borrowing current-state semantics.
• Make historical maps inert:
  • no green current-work dots
  • no active pulsation
  • no moving flow dashes
  • no follow affordance
• Reserve playback controls for History/review and dock them beneath the map so they cannot obscure graph content.
• Add 5x and 10x playback speeds alongside the existing review speeds.
• Preserve deterministic seek, play, pause, review-current-session, and Resume live behavior.

Agent Activity map and interaction design

• Standardize Claude Code, Codex, and OpenCode presentation around one visual grammar where the available evidence supports it.
• Distinguish process presence, meaningful current work, in-flight operations, and static historical relationships.
• Add long-running presence breathing and current-work animation without implying activity in History.
• Keep active flow dashes tied to a specific observed operation.
• Improve automatic layout to reduce overlap among actors, operations, labels, and callouts.
• Keep Fit, Focus, Reset, zoom, and Help controls in compact segmented groups.
• Support dragging and pinning individual actors and operation cards.
• Preserve pins and the user camera across live deltas.
• Reset either an individual placement or the full automatic layout.
• Add evidence-aware rollover descriptions and persistent selection details.
• Add a consultable Help/Legend covering actors, workers, operations, motion, Git counts, Live, and History semantics.
• Fix the Legend Close action and return keyboard focus to the Help trigger.
• Replace internal or synthetic labels such as <synthetic> with user-facing activity language.

Session Stream expansion behavior

• Add an upper-right chevron to collapse and restore Session Stream.
• Collapse to a narrow, persistent restore rail rather than removing the panel entirely.
• Let Agent Activity reclaim the released width on desktop.
• Use a compact collapsed row in stacked/tablet layouts.
• Preserve transcript and graph EventSource connections while collapsed.
• Preserve transcript selection, unread state, graph camera, and pinned positions.
• Persist the user preference locally.
• Add accessible labels, aria-expanded, live-region feedback, focus treatment, and reduced-motion behavior.

Session hierarchy and provider consistency

• Preserve native Codex worker-thread hierarchy from authoritative thread and spawn-edge evidence.
• Represent Claude sidechains as nested worker views without fabricating separate native sessions.
• Keep OpenCode within the common session shell while remaining honest about capabilities its source does not expose.
• Keep host, inference provider, and model as independent axes throughout projection and presentation.
• Use official OpenAI/Codex and OpenCode O marks, with authoritative adjacent text for accessibility and disambiguation.
• Remove the incorrect generic Codex target icon.

Live evidence, workspace context, and history

• Add native transcript discovery for supported live hosts.
• Capture safe workspace facts:
  • repository/workspace
  • working directory when meaningful
  • branch
  • tracked added/deleted line counts
  • changed file count
  • capture time and confidence
• Persist workspace captures for retained History inspection.
• Add workspace facts to selected-session details and transcript context.
• Remove redundant project names and unhelpful Directory: repo root text.
• Render branches with a branch SVG and concise branch name.
• Clarify that working-tree line counts describe the captured tree and are not automatically attributable to one session.
• Improve process/session correlation so stale or unrelated Codex sessions are not presented as recent activity for a Claude-only project.

Usage session identity and provenance

• Fix the malformed claudeunknown / codexunknown pills.
• Render the execution host as a clean, independent badge.
• Expand session details to show:
  • execution host
  • inference provider
  • model(s)
  • classification basis and confidence
  • token split
  • tools
  • flags/worktree
• Treat empty and literal unknown values as missing evidence rather than user-facing identity.
• Show Not recorded (not established by source) when a historical source did not establish the inference provider.
• Do not infer an inference provider from host or model names.
• Ingest Codex session_meta.model_provider as observed provider evidence.
• Bump the Usage index schema to v8 so cached Codex rows are safely re-derived.
• Retain OpenCode provider identity only when its native data supplies it.
• Keep Claude historical provider unknown when the transcript does not establish whether inference came from Anthropic, Bedrock, Vertex, Foundry, a gateway, or another endpoint.

Documentation and architecture

• Add a user-oriented dashboard interaction guide in docs/DASHBOARD.md.
• Replace docs/LIVE-SESSIONS.md with docs/OBSERVABILITY.md.
• Rename and update the Observability DDD document and context map.
• Rename and update ADR-0012 for Observability.
• Update ADR-0005, ADR-0009, ADR-0016, ADR-0021, and ADR indexes as living plans with current status/date notes.
• Document:
  • the three-area information architecture
  • canonical routes
  • Live versus History semantics
  • Session Stream collapse behavior
  • host/provider/model provenance rules
  • Claude/Codex/OpenCode capability differences
  • animation and historical-inertness rules
  • workspace capture and Git-count limitations
• Refresh transcript and scorecard source citations after the Usage schema change.

Quality and regression coverage

• Extend unit and integration coverage for:
  • provider provenance and Codex model_provider
  • thread-ledger and worker relationships
  • live adapters and projection behavior
  • workspace/Git capture persistence
  • Live/History separation
  • historical inertness
  • Help/Legend behavior
  • official provider marks
  • drag, pin, reset, fit, pan, zoom, and camera persistence
  • Session Stream collapse/restore and connection preservation
  • responsive layouts
  • 5x/10x playback
  • stable primary/secondary navigation
  • Usage identity expansion and missing-evidence language
• Validation completed locally:
  • pnpm run check
  • pnpm run test:ui
  • 193 browser/UI checks passed
  • no console errors
  • no failed browser requests
  • git diff --check