fix(runtime): adopt the Sandbox run result

[drewstone]

Summary

• Use the Sandbox run outcome tracker as the only terminal result source.
• Project Sandbox messages, tool calls, usage, and events once inside Runtime.
• Preserve blocked, failed, and completed run states after stream recovery.
• Remove the older event-level failure parser.

Why

Consumers should receive one complete Runtime result from the public Sandbox contract.
Braid no longer needs a second parser for raw Sandbox events.
A recovered tool error does not incorrectly fail a completed turn.

Dependency cohort and release

Sandbox 0.31.0 is published and exports createAgentRunOutcomeTracker and AgentRunOutcome from its runtime entry, so the dependency update is part of this pull request.

• Sandbox catalog and peer range become >=0.31.0 <0.32.0. The floor is 0.31.0 because this code requires createAgentRunOutcomeTracker, which 0.29 and 0.30 do not export.
• Interface catalog and peer range become ^1.3.0, because Sandbox 0.31.0 requires Interface ^1.3.0. One consumer install therefore resolves one Interface copy.
• scripts/lib/packed-package-test.mjs states the one supported Sandbox cohort (0.31.0), and the packed-cohort consumer install proves it.
• ci.yml and publish.yml check out the Interface cohort source at @tangle-network/agent-interface@1.3.0 (a3c31b98).
• Runtime moves to 0.143.0. The bump is a minor one because ./kernel removes assertSandboxEventSucceeded and sandboxEventFailure. It also adds SandboxExecutorToolCall, which names one retained tool call on SandboxLeafOut.
• Bench moves to 0.8.19 with the same cohort.

Proof

Local, on the merge head:

• pnpm run lint: 607 files checked, 0 diagnostics.
• pnpm run typecheck: clean (tsc --noEmit plus the examples project).
• pnpm run build: 151 files, 9.29 MB, complete.
• pnpm test: 2,780 passed, 6 skipped, 105 failed. Every failure is a local macOS artifact and reproduces unchanged on origin/main: the same 20 files fail on both, and a per-test diff of the two runs lists zero failures unique to this branch. They are git-worktree and process-spawn timeouts in tests/candidate-execution-*, tests/mcp/*-harness, and tests/version-bump-check.
• pnpm run check:version-bump: the minor bump is paid by 0.142.3 -> 0.143.0, and the bench patch by 0.8.18 -> 0.8.19.
• pnpm run verify:package: 2,087 exports across 17 entry points, record current; static-import, publint, attw, and edge tool loop pass.
• pnpm run docs:check: API reference regenerated with no diff, and the freshness gate passes on the version pin, 3 substrate peers, and 10 citations.
• pnpm run verify:primeintellect: pass.
• pnpm install --frozen-lockfile: clean after the lockfile regeneration.

pnpm run verify:bench fails locally only where its tests read /proc/<pid>/stat and resolve jail containment, which Linux provides and macOS does not. The agent-bench CI job runs the same command on Linux and passes.

Simplification

Simplification: the event-level failure parser is deleted with its tests (sandboxEventFailure, assertSandboxEventSucceeded, describeSandboxError, TERMINAL_EVENT_TYPES), so the four Sandbox paths that used to each apply their own per-event failure rule now read one terminal state from the platform tracker; TERMINAL_FAILURE is left scoped to one tool call only.
Net: +814 / -379 lines across 44 files (src and tests: +667 / -259 across 29 files); 103 lines of parser and its test block removed, 2 public exports dropped, 4 terminal-truth call sites collapsed to 1 owner.
Tests: +2 cases plus 5 table rows in the streamAgentTurn: Sandbox outcome contract block (a failed outcome after stream recovery stays failed with its usage; blocked_on_approval / awaiting_question / awaiting_plan_decision settle as blocked, not completed; a recovered tool error does not fail a completed turn), -7 deleted with the parser they tested (4 sandboxEventFailure decoder cases, 2 per-event assert cases, 1 renamed). No test asserts the new range, the version, or a removed export — check:version-bump, check:api-surface, and the packed-cohort job prove those.
Not done here: #912's remainder — served backend identity on settle records, and an explicit empty-vs-absent output marker beside content. That is a separate PR in this program, tracked on #912.

Closes #781

[tangletools]

✅ Auto-approved drewstone PR — a14f3115

This PR was opened by the trusted drewstone account.

This approval is provisional and was applied by the local stand-in because the pr-reviewer webhook host is unreachable (2026-08-21). CI on this head is fully green. The full PR reviewer audit re-runs via the resweep when the service returns and will publish findings if it detects issues.