docs: make the specs stop lying, and number three invisible items

[codeitlikemiley]

Phase 1 of the approved plan. A survey of the tree against the docs found that this repo's debt lives almost entirely in prose — there is not one TODO, FIXME, unimplemented! or todo!() in crates/ — but several documents contradict each other or the code.

One of them misled this session into repeating "nothing is laptop-buildable any more", which was false.

Documents that disagreed with the code

	was	now
docs/README.md (read-order item 0)	"the blueprint and the seed", crates "seeded with core types and traits"	matches the root README: specification and implementation
docs/23 Phase 1 contents	counts "event store PG" as delivered	no PG EventStore exists — implementors are Memory/Jsonl/Sqlite, panday-harnessd holds an Arc<MemoryStore>, and session_events is M18.6's sync sink
docs/25 testing contract	states a credential_id conservation property as enforced	UsageRecord has no such field; marked pending
docs/16	marketplace in phase 4	phase 6, matching docs/23 and GOAL.md

"Postgres is M3.5" — wrong in seven places

M3.5 is ledger-rebuild-from-log; the PG lane was M2.3. Both are ✅, so a reader chasing the pointer concluded the work had shipped. Four were in docs; grepping after fixing those turned up three more in code comments.

Two were doubly stale: docs/20 and tenancy.rs both claimed "there is no SQL to lint yet" when the lint now walks 49 sqlx::query sites and 8 migrations — underselling the milestone rather than overselling it.

Three items given numbers: 104 → 107, and no work was added

These already existed and were invisible to anyone reading the milestone list:

• M0.2 was being cited in docs/23, with a table of Phase 2 exit clauses under it, but no bullet ever existed — the tracker was uncountable and unmarkable. Exactly the failure CLAUDE.md §4 warns about, hiding behind a number that looked real.
• M11.10 PG-backed exact cache — specified since docs/11 was written, owned by nobody.
• M14.8 egress proxy — without it a plugin's net: [host] is an all-or-nothing switch, not an allowlist. Safe today under --unshare-net, but the manifest field reads like a promise.

Both count lines now explain the increase, so it doesn't read as scope creep.

Process rules from this session's failures (handover §1.2)

• Check gh with gh api user -q .login, not by parsing gh auth status — the active account drifts to hexuria, which cannot write here, and status-parsing gave the wrong answer twice.
• cfg-gated code cannot be verified locally (ring needs a C cross-toolchain); plan a CI round trip.
• The integration lane is not in the local gate.
• Never pkill -f "cargo test" — it is not scoped to this repo.

Local gate

fmt ✅ · clippy --workspace --all-targets -D warnings ✅ · panday-gateway ✅ · panday-sdk ✅ (four crates had doc comments edited). mdbook isn't installed here, so the docs CI job is the authority on the book building.

https://claude.ai/code/session_017kFpYDqvz6sKGSkM4YKaRf