fix(coding-agents): exclude external facts from knowledge pages

[Sanderhoff-alt]

Summary

Knowledge Pages can currently summarize a dependency's implementation or decisions as if they belonged to the repository being worked on. This happens because document attribution records where a fact was captured, but does not identify what project the fact is about. This PR adds an explicit external-knowledge classification and prevents those facts from entering the repository's standard Knowledge Pages.

Fixes #3476

What changed

The coding-agents plugin adds knowledge:external to the entity-label vocabulary. The extraction guidance says to use it only when the fact is about another project's own implementation, configuration, or decision. A decision about how the current repository uses or integrates a dependency remains normal project knowledge. The external label is explicitly exclusive with the other knowledge:* categories.

Each seeded Knowledge Page now refreshes from its normal category tag and excludes knowledge:external with a deterministic compound tag filter. The exclusion is therefore enforced by the refresh query instead of relying on wording in the page's natural-language source query.

Existing pages are upgraded in place. On startup the plugin reads the Knowledge Base tree, matches the fixed root pages by name, and patches a page only when its trigger is missing or different. Page IDs, existing content, and synthesized history are preserved. New pages receive the trigger when they are created, and repeated startup runs are idempotent.

The Knowledge Base API now exposes the page trigger in tree responses and accepts an optional trigger field in node PATCH requests. The route forwards that field to the existing backing mental-model update logic. The field is optional, so older clients and older page behavior remain compatible.

The CLI now exposes the same capability as knowledge-base update BANK NODE --trigger <JSON>. The argument accepts the complete trigger object, including nested tag_groups, and is included in the CLI OpenAPI coverage check.

The Rust SDK does not commit generated source files. Its build.rs generates the models from hindsight-docs/static/openapi.json at build time, so KnowledgeNode.trigger, UpdateNodeRequest.trigger, and nested trigger tag groups are available after rebuilding. The service's native MCP has no Knowledge Base CRUD operation, so it has no corresponding parameter to add. The coding-agents MCP also keeps page CRUD internal to the plugin.

The OpenAPI specification and generated Python, TypeScript, and Go clients are updated together with the control-plane types and wrappers.

Compatibility

The deprecated per-harness plugins continue to use their existing mental-model endpoints and are not changed by this PR. They remain operational, but they do not automatically gain the new external classification. The current unified plugin upgrades pages created by earlier versions of that same plugin when the server exposes the trigger field. Historical facts are not reclassified automatically; newly extracted facts use the new label.

When an older server does not expose the Knowledge Base surface, the plugin treats 404, 405, and 501 responses as an unavailable optional capability. It skips page creation or upgrade and continues with bank configuration and memory operations. Older servers that return a tree without trigger are not sent a trigger PATCH, so their existing page behavior is preserved.

Verification

The coding-agents test suite passed with 504 tests, including 31 Knowledge Page tests. The CLI trigger parsing test passed, the CLI compiled successfully, and the OpenAPI coverage check reported all 89 operations and 115 request parameters covered. Client and documentation generation, repository lint hooks, and whitespace checks passed. The server Knowledge Base integration tests were not run because this temporary worktree does not include pg0-embedded, so its test database cannot start.