perf(mcp): compact unknown capability run responses

[yuelgrace1810-ops]

Problem

A saved tool-first evaluation guessed a nonexistent math.run capability. The resulting UNKNOWN_CAPABILITY text repeated every installed capability ID and consumed 12,664 model-visible bytes before giving the normal recovery hint.

The equivalent math.find failure already keeps the full catalog in typed structured_content while projecting a compact text error.

Change

• omit available_capability_ids only from the agent-facing math.run text projection for UNKNOWN_CAPABILITY;
• preserve the complete typed result, diagnostics, assurance, and recovery hint;
• add an MCP boundary regression proving the compact text stays below 2 KiB while the structured result retains the installed IDs.

On the observed trajectory, the bound guarantees at least 10,616 fewer visible bytes (83.8%).

Overlap and independence

This is independent of:

• fix(polynomials): improve Nullstellensatz artifact diagnostics #830, which changes Nullstellensatz artifact diagnostics;
• fix(eval): preserve itemless Codex usage telemetry #832, which fixes evaluation token accounting;
• docs: clarify exact Jacobian contract inspection #786, which clarifies exact contract inspection;
• the selective-router experiment running separately.

No mathematical operation, ranking, checker, scope, or assurance behavior changes.

Validation

After updating the branch to current main (including #834):

• focused MCP boundary regression: passed;
• full unit lane: 750 passed;
• full lint/format/complexity/dependency/dead-code lane: passed;
• component lane: 740 passed, 3 platform skips;
• MCP lane: 38 passed; the one loopback-bind test blocked by the sandbox passed when rerun with loopback permission;
• end-to-end lane: 8 passed, 1 expected Lean skip;
• build: source distribution and wheel passed;
• git diff --check: passed.