docs(agents): drop the fact-placement convention from AGENTS.md

[lesnik512]

Why

AGENTS.md carried a per-repo copy of a doc-placement convention: a four-homes table, an admission check to run "before writing a line anywhere", an ADR row prescribing NNNN-slug.md and revisit triggers, and a ratchet rule.

Doc placement is a domain-modeling convention. Restating it in every repo means ~25 copies that can drift from it, in a file loaded on every turn — the failure the removed text itself named: "it is always loaded, so a line that restates a docstring… costs every turn and rots in two places at once."

The local ADR rule was also weaker than the convention it duplicated: it had no reversibility gate, so any non-obvious rejected alternative qualified.

Piloted and merged in lite-bootstrap (#188); this applies the same cut here.

Design

Removed:

• the Where a fact goes section in full — four-homes table, the "before writing a line anywhere" admission check, "prose about mechanism has no home", the ratchet rule
• Workflow's "the spec for a change is its PR body" rule, the "two things outlive the PR" ADR sentence, "there is no change file and no lane to choose", and "there is no separate truth-home directory"
• Architecture's "behaviour detail has no prose home… run the admission check" tail, where present

Kept:

• the INVARIANT: docstring shape, with its ADR and lychee clauses trimmed
• one Workflow fact: unscheduled work becomes a GitHub issue
• every other section, unchanged
• inline docs/adr/ citations inside architecture prose. Those cite a specific decision at the point it is relevant; they are not the convention being dropped.

No replacement pointer is added, by design.

Non-goals

• No ADR is changed, removed or renumbered. docs/adr/ is untouched.
• No code or test changes.

Verification

The remaining docs/adr/ mentions in this file, if any, are inline citations to specific decisions, not a rule about when to write one.

[github-actions]

Benchmark gate

✅ gate passed

scenario	msg/s	delete/msg	WALrec/msg	WALB/msg	fpi	upd	del	dead_tup
consumer/w1/b10	889	1.000	6.80	918	242	5000	5000	10000
consumer/w1/b100	874	1.000	6.84	920	243	5000	5000	10000
consumer/w2/b10	1036	1.000	6.79	959	242	5000	5000	10000
consumer/w2/b100	1264	1.000	6.78	994	263	5000	5000	10000
consumer/w4/b10	955	1.000	6.67	921	243	5000	5000	10000
consumer/w4/b100	1450	1.000	6.86	1035	243	5000	5000	10000
consumer/w1/b100/tfbs100	5183	0.010	6.07	1116	243	5000	5000	10000
producer/w1/b100	2154	0.000	3.04	584	0	0	0	0

Gated (fails the build): delete_calls + tuple counters (upd/del/ins) + the producer's insert_calls, exact; select_calls within +2; wal_records within a 10% band. msg/s, WAL bytes and total calls are informational (timing/FPI noise).