fix(examples): end every offline example on a terminal frame

[drewstone]

Problem

Every offline example that scripts a worker has been settling as a failed run. Sandbox 0.31 decides a run's outcome from done alone — result carries the payload and is explicitly not terminal:

// @tangle-network/sandbox 0.31.0, runtime-api, observeEvent
if (type === "result") observeOutcomeFrame(state, type, data, false);   // terminal: false
if (type === "done")   observeOutcomeFrame(state, type, data, true);    // terminal: true
...
if (!state.terminalReached) return failed("Agent stream ended without a terminal event", ...)

Four fixtures emit result and stop. The output still parses, so the example prints numbers and looks alive — and every iteration carries sandboxOutcome: { success: false, error: 'Agent stream ended without a terminal event' }, so no verdict is produced and the loop reports the wrong decision.

Measured on origin/main before this change, against what each README documents:

example	README says	actually printed
quickstart/quickstart	decision: pick-winner — winner: shot 1	decision: fail — winner: shot undefined, all three shots rejected
driver-loop	decision: pick-winner / winner: shot 1	decision: fail, all three shots rejected
researcher-loop	the clean candidate wins, two example-tenant items	no winner — every iteration failed validation
quickstart/minimal	decision: done — 1 iteration(s)	same line, on a failed outcome

Nothing caught it: examples are typechecked, never run.

researcher-loop was broken a second way, and hidden a second way. It imported researcherProfile, ResearchTask and ResearchOutput from @tangle-network/agent-knowledge/profiles — a subpath that package publishes in neither 8.0.10 nor 10.7.0 (its exports are ., ./viz, ./cli, ./memory, ./sources, ./benchmarks). Those symbols live in this repo, at src/profiles/researcher.ts, published as @tangle-network/agent-runtime/profiles and documented at docs/api/profiles.md:1192. The example typechecked green only because tsconfig.examples.json excluded the directory.

Change

• The four fixtures append { type: 'done', data: { outcome: { type: 'completed' } } } — the frame the repo's own kernel tests use, and the one examples/agentic-data-creation already had.
• researcher-loop imports the preset from @tangle-network/agent-runtime/profiles, and supplies the caller-owned AgentProfile that researcherProfile({ profile, task }) requires. researcherProfile is not reintroduced upstream; the dependency stays pointing down.
• tsconfig.examples.json no longer excludes examples/researcher-loop, so pnpm run typecheck covers it like every other example.
• The two documentation claims that pointed at the wrong package are corrected: the example's README no longer tells a reader to pnpm add -D @tangle-network/agent-knowledge, and examples/README.md's row no longer describes it as "uses the optional agent-knowledge peer".

Proof

Every offline example, run after the change:

quickstart/minimal      decision: done — 1 iteration(s)
quickstart/quickstart   shot 0: reject … shot 1: PASS … decision: pick-winner — winner: shot 1
driver-loop             SHOT 0 [reject] → SHOT 1 [PASS] … decision: pick-winner, winner: shot 1
researcher-loop         decision: pick-winner, winner: iteration #0, score 0.700, valid true,
                        items (2) both [example-tenant], citations 2, proposedWrites 1
chat-handler            unchanged (it streams finalText through a different path)

Each now matches its own README line for line, including researcher-loop's documented proposedWrites: 1 — insert into example-tenant.

pnpm run typecheck            clean, and now covers examples/researcher-loop
pnpm run lint                 615 files, no fixes
pnpm run build                clean
pnpm run check:api-surface    2120 exports / 17 entry points, record current
pnpm run check:version-bump   consumer surface unchanged at 0.155.0 (examples are not published)
pnpm run docs:check           exit 0
pnpm test                     2803 passed / 173 failed across 22 files
  clean origin/main, same machine: 2797 passed / 171 failed across 21 files. The
  baseline's 21 files are a subset of these; the extra is tests/kernel/workspace.test.ts,
  a 20 s git-worktree timeout in the known macOS class. CI on Linux is the authority.

Simplification

Simplification: one exclusion deleted from tsconfig.examples.json, so the example set has one rule instead of one rule and an exception; the fixtures now all end the same way.
Net: +12 / -12 lines across 8 files; 1 typecheck exception removed, 1 dead package dependency removed from an example and its docs.
Not done here: examples are still not executed in CI, which is why four of them could report the wrong decision for as long as they have. Running them is a build-time cost decision and a separate change — I am not adding a gate on the strength of one incident. What this pull request does is make the failure visible the next time someone runs one.

Tests: +0, -0 deleted. The proof is the four runs above; a unit test asserting a fixture emits done would restate the fixture, and running the examples is the CI question named above.

Refs #954

[tangletools]

✅ Auto-approved drewstone PR — db08622e

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.