refactor(runtime): deduplicate backend fixtures in ai-sdk-backend.test.ts (scavenger T1 slice)

## Claim

Part of [**T1 — Runtime backend and kernel tests**](https://github.com/apache/maka/issues/4726) (scavenger sweep 2026-09), scoped to one file: `packages/runtime/src/__tests__/ai-sdk-backend.test.ts`. Cross-checked against open claims: this slice is unclaimed — #3909 decomposes the source file, not its tests.

## The redundancy (measured)

Baseline commit `2310035a3` (current `main`), file SHA-256 `9770d5ff…`, 16,608 lines:

| Metric | Count |
|---|---:|
| `createTestAiSdkBackend` call sites | 196 |
| Eligible object-literal calls (all six default keys present; no spread, duplicate, or computed keys) | 181 |
| Intentionally excluded calls | 15 |
| Removable repeated property lines | 993 |

181 calls repeat the same six fixture properties byte-for-byte (`sessionId: 'session-1'`, `header: header()`, `appendMessage: async () => {}`, `apiKey: 'sk-test'`, `newId: idGenerator()`, `now: monotonicClock()`). The 15 exclusions stay on the real constructor: 8 type-asserted inputs, 4 Codex checkpoint calls that deliberately omit `newId`/`now` (the backend's own defaults are the subject there), and 3 spread-based inputs whose override order is behavior.

## The plan (two PRs)

### PR 1 — fixture-only deduplication (up as #4946)

One suite-local `createBackend` wrapper (17 lines) supplies the six defaults; every eligible call passes only its scenario-specific input; a final spread keeps explicit overrides authoritative. Nothing else moves.

- **Measured result** (the sweep's acceptance shape: a fixture that existed in N places now exists once): **181 → 1**; net **−976 lines** (16,608 → 15,632); runtime leaf count **222 → 222** with byte-identical leaf-name lists; no test deleted; no assertion, test name, or scenario value changed.
- **Input preservation evidence:** every migrated input expanded with its omitted defaults was compared against the original input at the AST level — all **196** constructor inputs equivalent (per-call assertion over the whole file).
- Gates: build, full target suite (222 pass / 0 fail), typecheck, `biome format`/`lint` (no fixes needed), `git diff --check`.

### PR 2 — scenario consolidation (planned; lands after PR 1 merges)

1. The two stream-watchdog pause tests share one scenario row table and one body, remaining two independently registered tests with unchanged names, semantics, and result payloads.
2. The two manual-compaction availability tests become one case table of three independently registered tests, each omitting exactly one capability (context budget / durable writer / summarizer). Today's fixtures omit several capabilities at once, so `operation_unavailable` is reachable through any guard — the masks hide which branch a test proves. Present callbacks count invocations and throw distinct sentinels so a fabricated compaction cannot pass. This adds the currently missing no-summarizer case.

- **Measured result:** leaf count 222 → **223**; the three availability guards become individually assertable.

## Part B acceptance mapping

1. **Coverage may not fall.** PR 1: leaf-count delta 0 with identical leaf-name lists. PR 2: +1 test. No test is deleted anywhere.
2. **Every test touched must fail without the code it tests.** For PR 2's touched tests, five named mutants were validated during planning and will be re-run and recorded in that PR: removing the execution-time `pauseTarget?.pause()` or its matching `resume()` fails both watchdog tests; deleting `!policy`, `|| !recorder`, or `|| !summarizer` from the compaction availability condition fails exactly its corresponding case, with the other four tests passing. **Open question for PR 1:** it touches 181 tests' fixture lines but changes no assertion, and the AST-level equivalence above proves each test's input set is unchanged. Applying rule 2 per-test to a fixture-only migration would mean ~181 mutants; we propose family-grouped mutants (one production mutant per behavioral suite, recorded named failures) as a follow-up commit to PR 1 if that grouping is acceptable — raised here before review rather than after.
3. **Prefer driving the real entry point.** Unchanged: every test drives the same `backend.send` / `backend.compactHistory` / `runtimeExecute` entry points; the wrapper only replaces repeated fixture construction.

## Rules check

- One PR per coherent change ✓ — fixture dedup and scenario consolidation are independently reviewable and sequenced.
- No behaviour change ✓ — test-only; leaf-name diff and input equivalence above.
- Existing gates decide the argument ✓ — format, lint, typecheck, protocol epoch, ASF headers all pass locally.