feat(mcp): MCP server rebase, refactor, and seven new tools

[Wenzel]

Three logical units on one branch/PR, since GitHub allows only one open PR per
head branch and this one is still under review. Each unit is reviewable on
its own by its commit range.

Unit 1: rebase onto the unauthenticated master (commits 1-4)

Brings the MCP server branch onto post-Auth0 master as three commits instead
of ten, plus a docs-privacy fix. No new tools; answers "do the existing 10
still work with auth gone?" without new-feature noise.

1. feat: add paginated search sessions
2. feat: add MCP server exposing OS intelligence to AI agents
3. docs: keep the private product name out of the public repository
4. docs: document diff system semantics and MCP constraints

Release blocker fixed: searchWithSession carried
if (include_updates === true && !context.jwt) throw. context.jwt is now
always undefined, so that condition is always true: replaying it unchanged
would ship a search tool that throws on every call passing
include_updates: true. The gate is deleted rather than adapted.

Two pre-existing defects surfaced and fixed: the root npm test script
was bare jest while jest.config.ts is configured for ESM, so any suite
using top-level await import(...) silently failed to compile while the run
still reported success (tests/git-log/index.test.ts was never actually
running); and a regression guard for the "empty filter drops Blob leaves"
gotcha had been broken since the commit that introduced the fix.

Unit 2: tool manifest refactor (commits 5-7)

Structural work only, ahead of the new tools below.

5. refactor(mcp): add the uniform pagination envelope
6. refactor(mcp): extract ref and path resolution into resolve.ts
7. refactor(mcp): make server.ts a tool manifest

• New mcp/src/resolve.ts: the multi-hop (ref, path) -> hash chains the
  new tools share. Lives in the MCP layer rather than in new GraphQL
  resolvers: the API is a stable open-source artifact with a second consumer,
  and the hops are localhost-cheap.
• New mcp/src/pagination.ts: the { items, has_more, next_cursor }
  envelope, including a composite cursor for tools that page two connections
  at once.
• mcp/src/server.ts becomes a manifest that iterates tool modules and
  applies the content/error envelope once: 545 lines to 46.

The pre-existing 10-tool test suite passes unchanged: same names, same
descriptions, same schemas, same behavior, verified word-for-word against the
pre-refactor file during review.

Unit 3: seven new task-shaped tools, docs (commits 8-14)

8. feat(mcp): add GraphQL documents for the task-shaped read tools
9. feat(mcp): add get_struct
10. feat(mcp): add list_symbols and list_structs
11. feat(mcp): add list_tree
12. feat(mcp): add list_registry_key
13. feat(mcp): add git_log and get_commit_capabilities
14. test(mcp): pin the two silent-empty-result gotchas, point at the task-shaped tools
15. docs(mcp): document the 17-tool surface and add a README

Takes the surface from 10 tools to 17. The new tools take (ref, path, name)
and resolve the chain internally; the hash-based tools stay as an escape
hatch, now with a pointer line so an LLM doesn't default to whichever it saw
first.

Tool	Signature	Hops
list_tree	(ref, path, limit?, cursor?)	3
list_registry_key	(ref, hive, key_path, limit?, cursor?)	5
get_struct	(ref, blob_path, struct_name)	4
list_structs	(ref, blob_path, name?, limit?, cursor?)	3
list_symbols	(ref, blob_path, name?, limit?, cursor?)	3
git_log	(path, entity_type, commit_range, limit?, offset?, ...)	1
get_commit_capabilities	(ref)	1

Why task-shaped. With thin tools alone, "show me the _EPROCESS layout in
win11-24h2" costs five round trips and dumps a multi-thousand-row
getBlobsWithSymbols result into context to find one blob hash. get_struct
is the clearest case: reading a struct today means calling diff_nodes with
the same hash on both sides plus status_filter: ["UNCHANGED"]. git_log
closes the largest capability gap: "when did this change, and across which
update" is the question OSWatcher exists to answer.

Pagination. list_tree and list_registry_key page two connections at
once behind one opaque composite cursor; a connection already exhausted is
not requested again.

Two regression tests assert the outgoing request body, not the response:
the filesystem-diff filter and the max_depth: 1 rule, because both bugs
that already bit this branch fail by returning a plausible empty list that an
exception-only test would pass.

Overall verification

Unit tests per tool plus dedicated suites for resolve.ts and pagination.ts.
19 suites / 96 tests pass; build clean; no "grapheos" or em-dashes anywhere
under mcp/.

Live validation against the seeded corpus follows in a separate PR.

🤖 Generated with Claude Code

https://claude.ai/code/session_014WnrRvhmHcMPj6iwJ2NnzK