feat(web): categorized context usage display, budget surfacing, and lifecycle inspector [stack 1/2]

[ChrAlpha]

Summary

Bottom of a two-PR stack. This PR surfaces the fragment-first context orchestration data (#1085/#1099) in the web UI and makes the lifecycle store it reads from safe for long-horizon sessions; #1157 (Trajectory pane) stacks on it, replaces the Phase C inspector dialog with a DeepSeek-Harness-style trajectory, and adds the per-request traces and stored context texts. Review this PR on its own; nothing here depends on #1157.

• Phase A — categorized Context Usage panel. Revives the shelved composer context ring + session-info popover as a Claude-Code-style categorized display: the 17 ContextfragKinds + tool-def buckets map into 8 fixed-order categories (context-categories.ts), rendered as a segmented bar + legend (context-usage-breakdown.vue) with an estimate-basis header (~total / window (%)). Provider-reported input stays a separate fact row, and the provider-basis single bar remains the fallback for sessions without lifecycle data. The ring shows for native and ACP sessions, stays off for direct runtimes, and refreshes when a turn finishes (one invalidation owner, set-diffed across panes, plus one follow-up refetch once the lifecycle row has landed). It is keyboard reachable: Enter opens the popover with focus inside, the label speaks the usage percentage, Escape returns focus to the ring.
• Phase B — coherent denominator and compaction mark. GET /bots/:id/sessions/:sid/status returns the persisted budget_plan (window, output reserve, tool-def cost, estimator) and compaction (enabled, auto_tokens). The mark derives from the persisted plan window only, only for native runtimes, and only when a summarizer resolves through the trigger's candidate chain; the plan is omitted when the pane's model_id resolves to a different provider model than the newest snapshot. The UI uses budget_plan.window as denominator, draws the output reserve band and a single auto-compact threshold tick against the window (the trigger's measured quantity differs by turn path, so the tick states the threshold and makes no claim about which segment reaches it). Compact Now and the /compact slash command follow the same availability.
• Phase C — Context Lifecycle Inspector. A dialog over /context-lifecycle: newest-first collapsible turns (first expanded, the reader's expansions survive refreshes), status tone, composition bar, selection counts incl. trimmed, drop reasons with token cost, trust breakdown, window/stable-prefix/cache metrics, mutations, per-step drops, a prompt-diff tag per turn (first turn / system / tools / system+tools / history only), page-scoped aggregates that render only observed values, and has_more/legacy notices with a load-up-to-200 action. The inspector chunk loads on first open and its query unmounts with the dialog.

Long-horizon storage boundary

selection_decisions grows with the whole history on every turn, so a 1000-turn session made the list endpoint return 42 MB and /status detoast a ~1 MB JSONB per poll. This PR moves that cost out of every hot path:

• context_lifecycles.snapshot is now a bounded summary (about 3 KB per row); the per-fragment audit lives in a new selection_decisions column. Every write statement splits the incoming snapshot at the storage boundary (snapshot - 'selection_decisions' / -> 'selection_decisions') and uses COALESCE so a summary-only rewrite can never erase an existing audit. Writes return identity columns only, and reconciliation read-backs no longer project the audit.
• SelectionTrace gains trimmed and drop_reason_tokens, rolled up when the snapshot is built, so no reader needs the per-fragment audit. Assistant-message metadata copies persist the summary only.
• Migration 0146 adds the column, then splits and back-fills existing rows per team (it opens and restores the teams policy like 0120; it holds an exclusive lock on teams for its transaction, so run it with the server stopped as the compose stack does).
• /status reads one summary row, the list projects summaries, the per-run detail endpoint is gone, lifecycle queries never persist to localStorage, and a storage quota error pauses cache persistence for the page instead of retrying on every change.

Measured on a 1000-turn seeded session on the dev stack: /status 20 ms → 7 ms; /context-lifecycle?limit=50 42 MB / 463 ms → 297 KB / 12 ms (Postgres 1,271 shared buffers → 210); inspector open 0.7 to 4.5 s with long tasks → about 100 ms with none.

Trade-offs kept

• The audit column still grows with history on disk; retention (TTL or last-K runs per session) is a separate decision and is one UPDATE on that column.
• Runs whose run-table row was never written recover from the summary-only metadata copy and therefore carry no per-fragment audit.
• Pages beyond 200 turns need cursor pagination, not included here; steps/mutations per row are step-bound, not history-bound.
• After a manual Compact Now the display reflects the last completed turn until the next turn produces a snapshot.

Since opening

• Merged origin/main at 67fe0e6 (fix(compaction): tolerate slow summarizers and show live chat activity #1115, fix(contextview): never trim a compaction summary before newer raw rows #1116, fix(contextview): prune in-flight tool results before failing a protected overflow #1121, feat(compaction): adaptive failure backoff and actionable context budget errors #1122, feat(web): chat 过程展示两层化 —— Adaptive Title 分层 + composer 常驻运行条 #1145). Two conflicts: apps/web/src/store/chat-list.ts keeps both sides' exports (streamingSessionIds from this PR, isSessionCompacting / beginSessionCompaction from feat(compaction): adaptive failure backoff and actionable context budget errors #1122), and packages/sdk/src/index.ts was regenerated from the merged spec. Go build/vet/test and the web checks pass on the merge commit.
• The lifecycle inspector dialog stays in this PR as the smaller, self-contained surface; feat(agent,web): Trajectory pane with per-request traces, injected context texts, and compaction rows [stack 2/2] #1157 folds it into the Trajectory pane.

Test plan

• Unit (Go): fragment rollup and Summary(), handler composition, plan applicability, compaction mark derivation, legacy status fallback, compaction threshold policy.
• Unit (web): category mapping, view adapters incl. prompt diff, breakdown geometry, inspector turns, turn-end invalidation, cache persistence breaker, session context view.
• Postgres integration (go test -tags integration): query round-trip with the split, summary-only rewrites keep the audit, 0146 down/up round-trip on a pre-split row incl. the token rollup, migration as a non-superuser database-owner role.
• Live e2e (Playwright against the dev stack with a real provider turn): popover numbers reconcile with /status exactly, no selection_decisions on the wire, the inspector opens without per-turn detail requests, keyboard path ring → popover → inspector → back, turn end refetches once plus one follow-up.

QA

1. mise run dev, open a session with at least one assistant turn.
2. Hover the ring left of the model pill, or focus it and press Enter: categorized bar + legend, ~ totals, provider input row, output-reserve band and free space when the window is known.
3. Click Context Inspector: per-turn composition, selection and drop reasons, trust, metrics, prompt-diff tags; the dialog stays open while the popover closes; Escape returns focus to the ring.
4. Sessions without lifecycle data fall back to the single-bar display and an empty inspector.

⚠️ No human QA — this PR has not been verified by a human yet. Remove this line once a human confirms the happy path.