docs: fix the instructions that actively mislead their reader

[Destynova2]

Third pass. This one audited every doc against the audience it claims to serve: operator, contributor, coding agent. An agent cannot ask a follow-up question, so a wrong instruction is a defect rather than a nuance.

Instructions that were actively wrong

File	Said	Reality
AGENTS.md, CONTRIBUTING.md (x2), docs/how-to/contribute.md	gh pr merge --auto --merge	The repo disables merge commits. The command is rejected every time. I hit this myself earlier today.
llms.txt	5 links	Two files deleted earlier today, one path that never existed, plus the four phantom presets and "7+ core traits" against a traits.rs that now holds two.
AGENTS.md	"Disable features at compile time for smaller binaries"	Measured: 18.1 MB full, 16.6 MB without dlp/watch/etc. The flags gate the wiring in server/init.rs, not the code. Worth 8%, not a size lever.
docs/slices/MANIFEST.md	Hand-maintained LOC counts	media listed 1.8k while holding 5.6k; routing 3.9k against 6.3k; shared 1.1k against 2.2k.
CONTRACTS.md	[`DEFAULT_TENANT`](crate::storage::DEFAULT_TENANT)	A rustdoc intra-doc link inside a Markdown file. Can never resolve.

The manifest counts are now generated: scripts/sync-slice-manifest.sh, with --check for CI. A number that is wrong is worth less than no number.

Missing information that forced guessing

• AGENTS.md vs CLAUDE.md had no stated boundary and had drifted (AGENTS.md untouched since June 18, CLAUDE.md updated July 26, each holding unique content). They are not duplicates: AGENTS.md is the map (what exists, where), CLAUDE.md the manual (how to write and land code). AGENTS.md now says so and defers to CLAUDE.md on release mechanics instead of carrying a weaker copy.
• The crates.io namesake warning lived only in CLAUDE.md, though AGENTS.md is what an agent reads first. crates.io/crates/grob is an unrelated Windows buffer library. Duplicated where it prevents the mistake.
• Nothing said where tests go, while three conventions coexist (inline #[cfg(test)], sibling tests.rs, tests/). dlp/ uses two of them at once, which is intended but looked like drift. CLAUDE.md now states the rule the codebase already mostly follows, and marks tests/unit/ as legacy.
• Nothing documented how to add a provider. src/providers/README.md now traces it, and leads with the answer that matters: most backends need no code at all, just provider_type = "openai" and a base_url. When code is needed, the deepseek shape costs ~8 files and the gemini shape ~28, and the README says which is which.

Checked and deliberately left alone

• The 24 module READMEs. Both auditors independently called them the healthiest artifact in the repo. Accurate after the trait removal, with a consistent Purpose / Owns / Non-goals / Depends-on shape. Touching them would be churn.
• CONTRACTS.md. Referenced from CONTRIBUTING.md, intentional, and still true after the trait removal (it never named the deleted traits).
• Feature-flag gating. An auditor claimed dlp and tap are advertised but unenforced because features/mod.rs declares them ungated. Verified false: they gate the initialization path, so the behaviour really does disappear. Only the size claim was wrong, and that is now documented rather than "fixed" by ripping out working flags.

Verification

1727 tests pass; markdownlint clean; every relative link in llms.txt, AGENTS.md, CLAUDE.md, CONTRACTS.md resolves; sync-slice-manifest.sh --check passes; shellcheck clean on the new script.