G15: add design-before-edit task memory

[iantocristian]

Summary

• add bounded hybrid/structural task localization and an evidence-backed design schema
• persist task-scoped design artifacts with reuse, freshness, immutable refresh, and explicit brief activation
• expose jscout scout design, MCP design_task / implementation_brief, telemetry, and agent guidance
• keep designs out of broad search, overview, and semantic embeddings while retaining exact-ID access

Validation

• cargo clippy --all-targets -- -D warnings
• cargo test --all-targets (325 passed)
• npm test (56 passed)

All model-path tests use the fake gateway; no paid model calls were made.

[iantocristian]

I'm challenging this as an unnecessary product surface expansion

[iantocristian]

Sol rerun after eval/product fixes

Run: g15-010-sol (gpt-5.6-sol, high reasoning, memory+embed+checker+workflows DB)

Infrastructure/result integrity

• The MCP child received the explicit pi-ai gateway selector.
• Default origins retained both repository and workspace.
• The external Playwright browser server worked.
• Registered Next process teardown completed without the prior teardown timeout.
• design_task received the original task verbatim, completed once, wrote design 426, and implementation_brief was read before editing.
• The agent made one later incremental jscout search while investigating catch-all parameter encoding.
• The first-call reproduction with context_bytes=48000 now plans one 30,859-byte request, preserving both explicit seeds and pruning 24 candidates to the two mandatory declarations.

Product verification

• 327 Rust tests passed.
• 59 script tests passed.
• Sol's own checks passed: Prettier, ESLint, types, Next build, and 15/15 focused browser tests.

Hidden grade: failed

• Layer 1: fail (1 failed, 10 passed).
• Failure: shape-preserving prefetch rewrite rendered category: "promo" instead of the corrected "sale".
• Gold coverage: 1/7 files patched; only optimistic-routes.ts matched.
• The generated design selected the incompatible-dynamic-child/parallel-branch collision mechanism. Sol later added catch-all parameter mismatch detection, but it still did not implement the stale optimistic-prediction invalidation/retry path across cache.ts and the other gold propagation sites.

Usage

• Duration: 1,014,718 ms (~16m55s)
• Input: 9,081,905 tokens (8,878,080 cached; 203,825 non-cached)
• Output: 29,977 tokens
• jscout calls: 7 product calls — 3 search, 1 overview, 1 semantic memory, 1 design, 1 brief
• Command calls: 40 (5 failed during test-driven iteration)

Conclusion: the #45 workflow and eval mechanics now operate end-to-end, but this run does not validate G15's hypothesis-generation value. The design phase produced a coherent adjacent mechanism, not the mechanism required by the untouched hidden oracle.