Search relevance: add per-file diversity caps and source/test path weighting

[willwashburn]

Problem

relaywash__Search currently ranks mostly by match count, mtime, or path depth. A single large file can dominate results, and test or fixture files can crowd out source files even when the task is implementation-focused. This increases follow-up calls and makes the agent inspect less useful snippets.

Goal

Improve default Search result quality with lightweight ranking rules that do not require a persistent index.

Scope

Add ranking and trimming behavior for:

• Per-file result caps.
• Test file penalties.
• Fixture and generated-file penalties.
• Source file boosts.
• Spillover behavior so capped files can still fill empty slots if there are not enough diverse results.

Proposed defaults

• maxPerFile: default to maxResults / 3, minimum 3.
• Penalize paths containing common fixture markers such as fixtures, testdata, __fixtures__.
• Penalize tests by default when non-test source hits exist.
• Do not hide tests completely, because test references are useful for behavior discovery.

API shape

Add optional Search args:

{
  "maxPerFile": 5,
  "includeTests": "auto"
}

includeTests can be auto, always, or never.

Acceptance criteria

• Default Search results include more than one file when matches exist across multiple files.
• Test and fixture files rank below source files for default implementation searches.
• Explicit path filters still work as expected.
• Add unit tests with a large matching file plus several smaller relevant files.
• Add observation fields so the learning layer can see when diversity caps were applied.

Implementation notes

Likely files:

• crates/wash/src/tools/search.rs
• crates/wash/src/search.rs
• crates/wash/src/profile.rs
• crates/wash/src/hooks/post_tool_observe.rs

Keep schema additions stable and avoid dynamic descriptions.