fix(codemode): reuse tool discovery catalogs

[thdxr]

Issue for this PR

N/A — memory profiling follow-up.

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Reduce repeated tool-discovery allocation:

• Reuse one rendered catalog per Location while the registry revision and visible tools match. Registration changes invalidate it; executable definitions stay request-local.
• Build the catalog/search index on demand, so known tool calls skip discovery work.
• Render signatures only for returned search matches, share descriptions between catalog and search, and join final signature fragments to avoid retaining intermediate string ropes.

Only three production files are included.

How did you verify your code works?

After rebasing onto the latest Code Mode hooks refactor:

• bun run check
• All 1,267 Code Mode tests
• 368 focused Core registry, catalog, MCP, and session-runner tests
• Independent background review found no actionable production regressions

A local compiled-server stress benchmark (median of three pairs, Bun --smol in both) used two Locations, four concurrent sessions, 400 schema-heavy fixture tools, 32 prompts, 64 model steps, 32 successful tool calls, and eight searches:

Metric	Before	After
Peak RSS	1,388 MiB	642 MiB
RSS at final prompt batch	1,165 MiB	488 MiB
RSS after three seconds idle	828 MiB	488 MiB
Workload time	17.9 s	5.1 s

The deterministic localhost model and constant-result tool emphasize server processing overhead. These are synthetic stress-workload measurements, not an end-to-end inference speedup claim.

Screenshots / recordings

N/A — runtime allocation changes.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

from OpenCode