chore: migrate off the planning/ convention

[lesnik512]

Migrates faststream-concurrent-aiokafka off the planning/ convention and onto PR-body-as-spec, with CONTEXT.md owning the vocabulary and docs/adr/ holding rejected alternatives. Part of modern-python/.github#67.

Replaces planning/ and architecture/ with CONTEXT.md, docs/adr/, and
invariant tests, matching the layout rolled out in modern-python/.github#50 and
across modern-di. Tracking issue: modern-python/.github#67.

Nothing was lost with the deleted directories:

• planning/releases/*.md are the published GitHub Release bodies. Checked all
  seven against gh release view: 0.6.2-0.6.5 are byte-identical but for a
  trailing newline, 0.6.0 differs only in a link to a pre-rename spec path. The
  Releases are the record. 0.6.1 is the one exception and is drafted as an issue
  rather than fixed here: its published body was hand-condensed at publish time
  and ends with a "Full notes" link to blob/main/planning/releases/0.6.1.md,
  which 404s once this lands. The content survives at the 0.6.1 tag; the fix is
  a one-off Release edit, not a repo change.
• planning/changes/* are design and plan documents for shipped work. Git
  history is the record; the two rejected alternatives in them that outlive their
  PR were rescued (below).
• planning/deferred.md held one open item, now drafted as an issue.
• planning/_templates/, planning/index.py, planning/.convention-version,
  planning/audits/, planning/retros/ are convention machinery with nothing
  left to serve.
• architecture/README.md describes the promotion rule the convention imposed.
  The five capability pages are prose about mechanism that the modules state more
  precisely — this package's source is unusually densely commented, and every
  claim about dispatch, routing, prefix extraction, watermarks and the direct-ack
  guards was already pinned by a named test. The two claims that were not are now
  tests.

Decisions. The three records become docs/adr/0001-0003, frontmatter stripped
and revisit triggers kept: batch-subscribers-unsupported (0001),
health-delegation-chain-not-deepened (0002), control-signals-not-honoured (0003).
Two more are rescued from change files:

• docs/adr/0004 records that refusing a non-MANUAL subscriber was built and
  rejected — it breaks the broker-level registration contract by turning a
  legitimate mixed-subscriber app into a hard error on its ACK subscribers.
  From the guard-direct-ack change file. 0001 now scopes its refusal against it.
• docs/adr/0005 records that a run()-owning async CommitLoop was rejected
  twice: the wait-tasks carry no invariant, and moving them would force every
  scheduler test back onto the event loop. From the pending-commits and
  commit-scheduler change files.

Invariants. Two architecture claims were enforceable and unpinned:

• tests/test_rebalance.py::test_the_rebalance_flush_default_stays_under_aiokafkas_max_poll_interval
  reads aiokafka's own max_poll_interval_ms default rather than hardcoding
  300 s, so it fires on a dependency bump as well as on our own default. Verified
  red at the boundary (flush default raised to 300.0) and red on a simulated
  upstream drop to 5 s; still green when the flush default is tightened to 5.0 and
  when the unrelated shutdown timeout is raised to 45.0.
• tests/test_commit_scheduler.py::test_the_commit_scheduler_reads_no_clock_and_touches_no_asyncio
  pins the seam ADR-0005 defends. Verified red three ways — import time plus a
  time.monotonic() call, a from asyncio import Event collaborator, and an
  async def on an existing method — and still green when a new synchronous
  method and a typing import are added. Both modules restored byte-identically
  and the suite re-run: 182 passed, 100% coverage.

CONTEXT.md is authored from scratch; there was no architecture/glossary.md.
FastStream and aiokafka own most of the vocabulary and it says so rather than
redefining broker, subscriber, consumer group, partition or offset. Six
terms are local. The _Avoid_ audit cut three of the four candidate entries
because the rejected synonym appears nowhere: no "skip" is used for a
pass-through, no "ignore" for a route, and "ready" never names the whole pending
list. A fourth was reversed — "handler" for the user's subscriber function is not
a straggler but the dominant usage, including inside the user-visible
_DIRECT_ACK_REASONS strings, so the entry now disambiguates instead of
legislating. The one surviving _Avoid_, "in-flight handler" for a user task, is
proven and its six stragglers are fixed: four README lines, a test docstring and a
test assertion message. None is a user-visible program string.

One source edit was forced. middleware.py's nack guard message carries a
GitHub URL to planning/decisions/2026-07-28-control-signals-not-honoured.md,
deliberately absolute so a PyPI user without a checkout can open it. That is a
user-visible string and it now points at docs/adr/0003-.... README's two links
to the same record were repointed with it.

AGENTS.md gains Workflow and Where-a-fact-goes, loses the planning lanes and the
architecture promotion rule, and sheds the tooling notes the justfile already
states and the invariant list the tests already enforce. justfile and lint-ci drop
index / check-planning; the offline lychee gate lands in _checks.yml
(modern-python/.github#66). context7.json drops its planning excludeFolders
entry. docs/agents/domain.md is deliberately not written: origin/main has no
docs/ and AGENTS.md had no "Agent skills" section, so authoring it alone would
be a half-rollout of a separate convention with nothing to link it from.

The link gate went from 4 errors to 0. Three lived only in planning/ and are
fixed by deleting it; the fourth was a dangling docs/adr/0004 link left by an
interrupted earlier attempt, fixed by writing the record it wanted.

Beyond the recipe, and the reason this is chore: and not docs: release.yml read
planning/releases/<tag>.md as both a mandatory gate for stable tags and the
Release body source, so deleting planning/ would have failed every future stable
release. It now matches modern-di/.github/workflows/release.yml verbatim
(generated notes, prerelease flag only), which dropped the same gate in 1ea74ee,
modern-python/modern-di#449. This retires the mandatory-curated-notes policy for
this repo; a release wanting prose is edited after the fact.

Deferred work is drafted, unopened, in
DRAFT-ISSUES-faststream-concurrent-aiokafka.md one directory up: the commit_all
flush-urgency loss on a transient-error re-queue, the partition-scoped rebalance
flush, and the 0.6.1 Release link.

Verification

• just lint-ci — clean (eof-fixer, ruff format, ruff check, ty).
• Test suite green at the repo's 100% coverage gate.
• Offline link gate (lychee --offline --no-progress '**/*.md') — 0 errors.
• No planning/, architecture/, check-planning, check-links or convention-version reference remains.

Note

release.yml previously read planning/releases/<tag>.md, both as a hard gate for stable tags and as the Release body source; deleting planning/ without changing it would have broken the next stable release. It now uses GitHub's generated notes, matching modern-di post-modern-python/modern-di#449. This retires the mandatory-curated-notes policy.