[do-not-mirror][do-not-merge] chore(fork): scale process cost to change risk - #10

What

Makes the fork process cost what the change risks, and reduces the manual steps in it to exactly
one: pressing "Rebase and merge".

The gates stop being rituals and become a check. New fork-only workflow
.github/workflows/fork-gates.yml runs forkctl check-commit plus the upstream-readiness proof
(forkctl mirror <slug> --sha HEAD --no-publish) on every PR into integration/3.0, deriving the
slug from the branch and the chore(fork): exemption from the commit subject. Nobody runs those
commands, and nobody pastes their output into a PR description any more.

Review stops being a separate blocking step. It was: an adversarial review by a non-author, run
as a fresh-context agent for Track A, posted as a PR review listing every probe including the ones
that found nothing, with each finding resolved in-thread before merge. It is now the merge decision
itself — forkctl track still says how hard to read the diff, PROCESS.md still holds the Track A
prompts, but there is no approval to collect, nothing to post, and nothing written down unless
something was actually found.

Three gates become two checks plus the button. Gate 1 is CI. Gate 2 is Fork gates. The old
Gate 2 (review) is gone as a gate; the old Gate 3 is renumbered.

Three over-strict guards relaxed, all the same bug — conflating "this branch is correct" with
"this branch is current". Each one forced every open PR to rebase and re-run its gates whenever
anything else landed:

• check-commit: "not rebased directly onto the base" was a failure, now a notice. "Rebase and
  merge" already replays the commit and refuses when it cannot.
• resolve_feature_commit: required the commit be parented on the tip of integration/3.0, now
  only that its parent is an ancestor of the stack.
• fork_state_problems: reported a required fork-only file as deleted when the branch simply
  predates it. Now excused only when the file is absent from the branch's own merge base — deleting
  one you did inherit is still caught.

Smaller surface. PR template 54 -> 31 lines, with both gate code blocks and the three gate
checkboxes removed. Section 4.3 inverted: CI runs the suites, you owe pre-commit plus
make rebuild for C changes. Section 8 reduced to five bullets and no commands. CONTRIBUTING.md
and PROCESS.md follow.

New forkctl mirror --no-publish proves the gate without moving mirror/<slug>, so CI cannot leave
mirror branches behind. forkctl delete-mirror is fixed to match: it treated an absent remote branch
as a failed push, so a local-only mirror — now the normal case — could not be cleaned up at all.

Why

Two separate pieces of feedback. The cron jobs had been failing since 29.10 on false alarms, and the
per-PR cost had grown to: 2 gate commands run by hand, 2 pasted output blocks, 4 local suites, 7
checkboxes, a review ceremony with a written probe list, and a forced rebase every time anything
else merged. Most of that is mechanical and was already implemented in forkctl — it was simply
being executed by a person (or, expensively, by an agent) instead of by CI.

The parts that actually need judgement — is this the right change, and can it break the sim — do not
get cheaper by being written down twice. They get cheaper by happening once, at the moment someone
decides to merge.

Notes

The workflow was run (section 4.8 requires this of any PR that adds one):

• Chore path, this PR: green in 31 s, correctly detecting chore(fork): and skipping the upstream
  half.
• Feature path, throwaway probe PR [probe] exercise the feature path of the Fork gates check #18: both steps executed and reported from one run
  (if: always()), slug and chore detection correct, each failure carrying its fix. The probe is
  two commits by construction — it sits on top of this unmerged chore commit — so it is a red
  feature path by design.
• A green feature path cannot be produced in CI until this commit is in the base, because the
  workflow only exists on branches carrying it. Proven locally instead, with this commit's
  forkctl against the real single-commit branch of ci: skip the test jobs on documentation-only pull requests #15:
  mirror ci-skip-doc-only --sha HEAD --no-publish -> OK, clean cherry-pick onto pure
  upstream/3.0, mirror diff one file, branch not moved.

Two bugs the probe found that a review would not have: actions/checkout with ref: <sha> leaves
HEAD detached and mirror refuses to run there (fixed with git switch -c gate-head), and the
parent-on-tip rule above.

Tested: full forkctl suite, pre-commit run --all-files, forkctl check-commit --chore.

Cleanup done alongside this PR (stale state, not part of the diff): deleted the folded
chore/forkctl branch on origin, four orphaned local mirror/* branches with no upstream PR, and
two pre-rebase local leftovers. Every remaining branch now maps to an open PR or is structural.

This is a chore(fork): PR. It is folded into the base commit by fixup and closed unmerged
(section 3.3) — never merged with the button.

Post-fold step, do not skip: add Fork gates to the required checks on integration/3.0. It is
deliberately not required yet — the workflow does not exist until this commit is in the base, and
a required check that cannot run sits pending forever and blocks every merge.

Suggested order (one history rewrite, one rebase per PR):

1. Merge fix(ci): skip the fused encoder where triton is absent #13 — it clears the triton failure that is red on every other PR.
2. Fold this PR, rebasing onto origin/3.0 in the same step (closes issue Rebase integration/3.0 onto upstream 3.0 #19).
3. Add Fork gates to the required checks.
4. Rebase ci: skip the test jobs on documentation-only pull requests #15, docs: align the comment standard across AGENTS.md and CLAUDE.md #17, fix(eval): resolve carla_cosim benchmarks and lint for unresolved names #14 onto the new tip; they pick up both the triton fix and the gate.