feat(desktop): show trustworthy turn results / 展示可信的本轮结果

[SivanCola]

The completion card previously mixed mutation counts with verification guidance and could imply that a turn had been verified without showing the commands that actually ran. This change makes it a read-only Turn results entry: confirmed net file changes, recorded check outcomes, and direct access to the evidence for that specific turn.

• Freeze checkpoint-owned net changes when the turn ends. Repeated edits, creates, deletes, no-ops, binary files and mode changes are handled explicitly. External changes, missing snapshots and size limits produce partial/unknown coverage; they never become an invented zero or an exhaustive workspace claim.
• Keep historical patches stable, even after later edits or rewind. The existing checkpoint store retains the optional result, and the existing display sidecar carries its summary. Retention expires patch payloads while preserving the summary. Processing and retained patches share a 2 MiB turn budget.
• Show running, passed, failed, interrupted, stale, unverified and unavailable check states from execution receipts. Details expose the original command, exit code and bounded log. Stable local message IDs prevent a reused provider call ID from opening another turn's log.
• Add View changes and View check details actions, a per-turn workspace panel, and a link back to the full workspace view. Session, turn, result and request identity guards reject delayed responses after navigation. These controls do not execute or retry commands.
• Preserve legacy history compatibility and the existing full/windowed transcript kernel. Add English, Simplified Chinese and Traditional Chinese copy, plus bilingual documentation.

Validation covers the real controller completion path (success/error/cancel), frozen snapshots and retention, event-wire round trips, stable log sources, delayed UI responses, history hydration, and current-versus-historical inspection. Browser checks use the actual Transcript, reducer, result panel and DiffView with fixture execution APIs; backend tests independently exercise real checkpoint/controller storage. Light/dark themes, 280 px layouts, long paths, failures and history selection were inspected in the browser.

Local implementation gates passed: root and Desktop go test ./...; go test -race for checkpoint/control/completion/eventwire and focused Desktop result/display paths; root and Desktop golangci-lint; repolint; frontend typecheck and all 330 standard test suites; production build and bundle budget; full macOS ARM64 wails build -clean including generated bindings, frontend, native compilation and packaging. CI exposed one additional dedicated test that still expected the old default workspace verification panel. The follow-up test-only commit fixes that assertion, and all 386 commands from node scripts/run-ci-tests.mjs now pass locally. GitHub checks are tracked on the published head separately.

Build size was measured against main-v2 at 45239b1: raw runtime JS 2,464,923 → 2,477,508 bytes (+12,585 / 0.51%); gzip runtime JS 454,722 → 458,209 bytes; CSS 122,869 → 123,257 bytes. Only the measured CSS, locale and raw-JS budgets receive narrowly rounded increases. The initial gzip-JS budget is unchanged. Source extractions keep the existing ownership boundaries; the only per-file line-baseline increases are WorkspacePanel +7, bridge +2 and types +8. Global repolint limits are unchanged.

Related work: merge-tree checks with #9911 (checkpoint retention configuration) and #9929 (transcript eviction) are clean; their later integration should retain result-payload accounting/expiry and result record sizing. #9966 requires coordinated semantic conflict resolution in check-bundle-budget.mjs, TranscriptCards.tsx, transcriptStore.ts, types.ts, useController.ts and eventwire/tool.go, preserving both its read status and this PR's result/Verifying fields. These PRs are not dependencies or included implementations. The current main-v2 settings-layout update at 7717f3e merges cleanly with this branch.

Limits: only changes with trustworthy checkpoint ownership are counted; historical data without receipts remains unavailable; log payloads may disappear with existing history retention. Bounded content processing does not establish a worst-case filesystem latency guarantee. Native Windows/Linux behavior remains subject to their CI jobs.

Documentation-impact: updated - document turn result coverage, frozen historical diffs, read-only inspection, recorded checks, log identity and retention in docs/TURN_RESULTS.md and docs/TURN_RESULTS.zh-CN.md, linked from the checkpoint guides.
Cache-impact: none - result metadata is host-owned receipt/display data; system prompts, provider request serialization, tool names and tool schemas are unchanged.
Cache-guard: existing provider-prefix and schema guards are exercised by the root Go suite; focused receipt/eventwire tests verify the added display metadata contract.