docs-consistency-check

Docs/metadata consistency

Check every one of these, not just a sample:

• README.md's tool table matches src/server.ts's registrations (names, and the keyless-vs-key-gated column against each tool's actual behavior).
• manifest.json's tools and prompts arrays list the same names as what's actually registered (server.json has no tool list; npm test's version.test.ts already asserts this — treat a failure there as authoritative). Their description fields are deliberately short, independent marketing-style summaries, NOT a copy of the tool's full .describe()/description text in src/tools/*.ts — don't "fix" them to match verbatim, that's not a bug. Do re-read them for accuracy if a tool's behavior changed in a way the short summary now misrepresents — manifest.json's user_config descriptions too (steam_id kept saying a vanity name needs the key after keyless resolution shipped).
• Tool description/field .describe() text in src/tools/*.ts itself: does it still match the actual inputSchema/outputSchema and the real behavior? Cross-check new/edited descriptions against the tool-description-check skill (Glama's TDQS rubric) per AGENTS.md. Check each outputSchema branch against what the shaper actually builds, not just the happy path — get_owned_games' found:false branch still advertised an owns field the 0.10.1 fix had stopped emitting, i.e. the published schema re-legitimized the very bug that was fixed.
• CHANGELOG.md's [Unreleased] section (see the changelog-style skill for entry style) has one line per real behavior change made in this pass — add missing entries, don't just flag them as missing.
• docs/architecture.md's "Keyless caveat" list still matches the actual keyless method list in src/clients/, and AGENTS.md's src//scripts/ tree (and this .agents/skills/ entry) still match the filesystem.
• docs/notes.md, docs/clients.md and any other docs/*.md for stale phrasing (e.g. describing something as "once published"/"upcoming" that already shipped).
• PRIVACY.md and SECURITY.md: grep src/clients/ for every host the code contacts and diff it against their host lists (keyless vanity resolution added steamcommunity.com while both still said "two hosts"). Then re-verify every other specific claim against the actual current code, don't just skim for plausibility — which credentials exist and how each is transmitted/redacted, what is and isn't cached (incl. cache key/TTL), the current list of read-only vs. write/mutating tools, and any host-allowlist statement. This class of drift is easy to miss because it reads fine on its own and only breaks against the code: this repo's own SECURITY.md/PRIVACY.md both claimed "player-specific data is never cached" after a later feature (get_player_summary's Steam-level cache) added exactly that, and a separate claim conflated an actually-cached field ("review histogram") with a similarly-named never-cached one ("reviews") — neither doc was self-evidently wrong, both required re-reading the client code to catch.