docs: move inline narrative out of the AGENTS.md rules that already have a decision record #928

Problem

AGENTS.md loads into every agent session, so its length is a cost paid on every
run. #921 (PR #925) took it from 9,677 to 8,603 words by rewriting prose, and
that lever is now spent — 50 of 96 rules were rewritten and further wordsmithing
would be scraping.

What remains is structural:

Over half the file is narrative that already has a home. The file's own
preamble says a → link means "the narrative is behind that link in
docs/decisions/", yet those 48 rules average ~98 words each and carry the
reasoning inline as well as linking to it. The records are substantial, not
stubs: 530-multi-farm-tenancy is 4,079 words, 782-ci-job-gating 2,012,
271-single-serving-instance 1,479, 846-adapter-reach-ratchet 1,473.

This is not a repeat of #413, which moved rationale into the records. This
removes what stayed behind afterwards.

Scope

For each of the 48 rules carrying a → link, answer one question and edit to
match the answer:

What is the minimum an agent must know to avoid breaking this rule, without
opening the record?

Keep that inline. Move the rest into the record the rule already links to.

What must stay inline — this is the entire risk

An agent who never opens the record must still be unable to break the rule.
These never move, however long they leave the paragraph:

• the imperative — what to do and what never to do, including the negative form
• every identifier: command, path, file, env var, config key, class, test name
• every exception and opt-out (Database:AllowInsecureConnection=true,
  RateLimiting:AllowNoTrustedProxies=true, the IdentityProvider exemption, …)
• every accepted risk and statement of what is not guaranteed —
  "Nothing enforces this.", "only review catches them", "neither survives a merge
  to main", "never fail-open". A rule that loses its limit reads as a
  stronger guarantee than the repo actually offers, which is worse than a
  long rule.
• any counter-example or wrong-fix warning that heads off a plausible mistake

What moves is incident narrative: which review round found it, what the earlier
broken version did, the measurement behind the decision.

Target

~6,300 words (−27%) if linked rules land near 50 words each. This is a
prediction, not a quota. A rule needing 150 words to stay safe keeps them, and
the PR names which and why. Going materially below ~6,300 means deleting rules,
which is a separate decision.

Acceptance criteria

• All 48 linked rules reviewed; the PR reports how many were shortened and
  how many deliberately left, with reasons for the latter
• Zero loss of decision links, issue references, runbook links, backticked
  identifiers, exceptions, accepted risks or negative prohibitions —
  verified by count before and after, not by reading
• Text removed from a rule lands in that rule's record; nothing is dropped
  outright without saying so in the PR
• Each edited record still reads coherently after the insert
• Every rule remains one paragraph
• PR records before/after word counts for AGENTS.md and for every record
  that received text
• No rule changes meaning, and docs: /deslop AGENTS.md without weakening its rules #921's "do not track issue state here" rule
  is respected
• PR body stays short: this repo squashes with the PR title and a blank
  body, and long accounting belongs in a PR comment

Non-goals

• Deleting rules or relaxing any policy
• Touching the 49 unlinked rules — with no record to move text into, shortening
  those is deletion, and they carry the least written history
• Creating new decision records
• Hitting a word target at the cost of an accepted risk

Known follow-ups already identified

Three code-versus-doc mismatches were found during #921 and deliberately left,
because each needs a precision edit rather than a deletion. Fix them here or
split them out:

• gh pr edit is documented as failing with a Projects-classic GraphQL
  deprecation error. It works — tested on PR feat(web): redesign Sales with a MUI order desk and settlement rail #927, exit 0, gh 2.101.0.
• Developer experience: add an Aspire AppHost for local orchestration and observability #565 says the AppHost is "deliberately not in CI"; ci.yml:189 runs
  - group: apphost as a matrix leg. The claim is really about aspire run's
  config wiring, not AppHost.Tests, so the wording needs narrowing.
• Sim harness (#243) rotted silently: 4 breakages, no CI ever ran it #370 says the simulation harness is "deliberately not in CI";
  e2e-smoke.yml and k6-baseline.yml both drive tools/simulation. The claim
  is really about the boot-guard harness (bootstrap.sh,
  docker-compose.sim.yml, verify-harness.sh), and Daily entry: require grading to reconcile sellable eggs before submit #394 in the same file
  already says the Playwright specs run on every PR — so the file currently
  contradicts itself.

Worklist

All 48 linked rules, heaviest inline first.