refactor(router): route simple intents to workers_ai as primary executor

[stackbilt-admin]

Closes #71

Changes

web/src/kernel/router.ts

• DEFAULT_ROUTES: greeting, general_knowledge, tarot_pulse, tarot_trajectory → workers_ai
• selectDefaultExecutor: greeting fixed route now returns workers_ai
• Verify zone (confidence 0.50–0.79): low-complexity no-tool → workers_ai (reverts the // workers_ai → gpt_oss bump)
• Degraded procedure fallback: DEFAULT_ROUTES[classification] ?? 'claude' → ?? 'workers_ai'

web/tests/router.test.ts

• 6 assertions updated to reflect the new routing targets
• 5 new tests: general_knowledge, tarot_pulse, tarot_trajectory, tarot_deep, greeting baseline

Guardrails preserved

• memory_recall stays on gpt_oss (equity recall failure 2026-03-04)
• code_task stays on claude_code
• Tool-requiring intents stay on gpt_oss
• Complex/high-complexity escalation to claude/claude_opus unchanged

Test result

118/118 passing (executor-router regression + semantic + router)

[stackbilt-admin]

APPROVED

The routing changes are correct and match the issue spec. Posting one structural note and one flag on the bundled scope.

Routing correctness (✓ all match spec)

1. DEFAULT_ROUTES — greeting, general_knowledge, tarot_pulse, tarot_trajectory → workers_ai (web/src/kernel/router.ts:102–113)
2. memory_recall stays on gpt_oss with the equity-recall comment (router.ts:103) ✓
3. Degraded fallback ?? 'claude' → ?? 'workers_ai' (router.ts:452) ✓
4. Verify-zone low-complexity no-tool → workers_ai (router.ts:172) ✓ — reverting the previous intentional bump back to where the issue spec wants it
5. selectDefaultExecutor greeting fixed-route updated to match (router.ts:137) ✓

Test coverage (✓)

• New tests: general_knowledge, tarot_pulse, tarot_trajectory, tarot_deep (counter-test keeping gpt_oss), greeting baseline (router.test.ts:1096–1118)
• Verify-zone test renamed + assertion updated (router.test.ts:1125, 1134)
• Degraded-procedure and broken-procedure tests updated (router.test.ts:1152, 1162)
• executor-router.test.ts untouched — correct, that file tests executor registration not dispatch routing

One structural note (non-blocking)

The internal trigger path at router.ts:219 and :240 still has ?? 'direct' as its unknown-classification fallback. The issue spec called out "line 452" specifically, so the internal path is arguably out of scope. Worth a follow-up issue if you want the internal path to also fall to workers_ai instead of direct for unknown classes — but it's a different semantic (internal triggers with unknown classifications are a different error condition than external routing misses).

Bundled scope flag

This PR bundles a substantial unrelated change: the graph.ts WasmGraph refactor (new @stackbilt/wasm-core dep, vite-plugin-wasm, vite-plugin-top-level-await, rewritten test fixtures across four test files). The WasmGraph WASM boundary introduces a test-correctness risk: the pruning test at recall-baseline.test.ts:708 now asserts expect(weak).toBeUndefined() against opaque WASM behavior — if WasmGraph.spreadActivation returns sub-threshold nodes, that assertion will fail with no clear signal. This isn't a blocker for the routing changes, but if the graph work ever needs to be rolled back independently, it can't be done cleanly. Consider whether this should have been a separate PR.