Delete the local PR templates; keep the generic org default #53

What changed

This issue was "host the PR-body-as-spec template as the org default, and retire
the local copies". Measuring the templates before sequencing the swap inverted the
conclusion: the PR template does no work in this org, and hosting it centrally
would not change that. The work is now three deletions and four reference fixes.

The measurement

Mechanism. gh pr create reads .github/PULL_REQUEST_TEMPLATE.md only in its
interactive editor path or when passed -T/--template. --body and --body-file
skip it. gh 2.98.0's own help is explicit: "--body to skip this". An agent
opening a PR with --body-file never touches the template. Only the web compare UI
applies it automatically.

Authorship. 26 of the 28 non-archived org repos have exactly one PR author,
lesnik512, opening PRs through gh. The two exceptions are
fastapi-sqlalchemy-template (4 outside humans across 47 PRs) and that-depends.

Where the convention checklist actually appears, in the four repos carrying it:

Where a template does work. that-depends is the only repo with real
contributor traffic: 9 distinct PR authors, lesnik512 a minority at 28 of 100,
xelandernt at 54. Across its last 25 PRs, xelandernt (web UI) filled the
template in 8 of 9, checklist included. lesnik512 (agent, --body) filled it
in 0 of 16.

Where a template does not work even on humans. fastapi-sqlalchemy-template
inherits the generic default, so its 4 outside contributors were all shown it.
Their PR bodies: 1726 characters, 159, 78, and one empty.

What did spread. The Why / Design / Non-goals / Verification shape is
used consistently, including in this repo (#62, #63, #65) whose own template says
Summary / Changes / Checklist. faststream-outbox #141-#157 carried the
convention template locally and produced bare ## Verification bodies. The shape
tracks AGENTS.md, which is loaded on every turn. It does not track the template,
which is never opened.

The conclusion: what the agent reads persists, and what it does not read decays. The
convention already lives in each repo's AGENTS.md, which is the only channel that
reaches the party writing the PR.

Why not host it centrally anyway

Two reasons the original plan does not survive the measurement.

It would duplicate AGENTS.md. The checklist's four routes restate the
"Where a fact goes" table and the two-things-outlive-the-PR paragraph that each
adopting repo's AGENTS.md already carries. By the admission check, a fact
derivable from a file the agent already loads does not get written.

AGENTS.md is not inheritable, so the central-hosting argument does not
transfer. GitHub's default community health files are CODE_OF_CONDUCT.md,
CONTRIBUTING.md, discussion category forms, FUNDING.yml, issue and pull request
templates plus config.yml, SECURITY.md, and SUPPORT.md. AGENTS.md is not on
that list, and no repo's AGENTS.md references this one. The 27 copies are
independent by necessity; that duplication is accepted deliberately rather than
solved with a generator.

The work

.github (this repo)

• .github/PULL_REQUEST_TEMPLATE.md: keep it as the org default, and delete the one
  checklist line Docs updated if behavior or public API changed. That line asks the
  pre-convention question ("did you update the page?") and is the only part of the
  generic form that is untrue for a repo on the convention. Everything else stands:
  lint, type check, tests, build, metadata are true under either convention.
• AGENTS.md: drop the opening claim that its conventions apply across all repos in
  the org. No other repo has ever read this file. Reword the Workflow blockquote so
  the generic default reads as a deliberate choice rather than a pending replacement.

modern-di, faststream-outbox, chat-app

• Delete .github/PULL_REQUEST_TEMPLATE.md. All three fall back to the org default.
• AGENTS.md: drop the sentence naming the template as the carrier of the shape. The
  AGENTS.md line "the spec for a change is its PR body: why, design, non-goals,
  verification" stays and is now the whole statement of it.

that-depends

• Nothing. Its template is the only one measurably in use, its mypy/pyrefly
  checklist line is a legitimate reason to stay local, and it keeps its own copy.

Deliberately not done

• The per-section guidance is dropped, not relocated. The templates were the only
  place explaining what belongs under each heading. The four section names carry
  enough on their own; moving ten lines of prose into four AGENTS.md files trades a
  file nobody reads for weight on a file loaded every turn.
• No ADR. The reasoning is this issue body.
• No drift check. With no local templates left outside that-depends, there is
  nothing to silently override, which was the only reason modern-di and
  faststream-outbox held their copies back.
• Nothing about the 22 planning/ repos. The original blocker was that a
  convention template would land on repos whose own AGENTS.md forbids it. With the
  generic default staying, nothing collides, so this issue neither needs nor waits on
  their migration. That migration is tracked independently in Migrate the remaining repos off the planning/ convention #67.
• The revisit-trigger clause is dropped. modern-di's template said an ADR needs
  one and faststream-outbox's did not; neither repo's AGENTS.md says it. It is
  not being added. chat-app enforces it mechanically in just check-adrs and is
  unaffected.

Revisit trigger

A repo other than that-depends starts taking regular outside contributions. At that
point a template becomes load-bearing for that repo, and it should get a local one
matching what its contributors need.

Related

• Host the PR-body-as-spec template org-wide; adopt the convention here #50, Rename CLAUDE.md to AGENTS.md org-wide, with CLAUDE.md as a symlink #51 (both closed; their step 1 and the AGENTS.md rename have landed)
• that-depends has no agent instructions file #69 — that-depends has no agent instructions file
• chat-app is half-migrated: the PR body is the spec, but planning/ is still a merge gate #70 — chat-app is mid-migration off planning/