[Feature]: Collapse Multi Merges Onto Main

[wallstop]

Feature Description

Right now something merges, then we trigger perf, then we trigger llm regen.

This should be collapsed into one flow

Use Case / Motivation

Too many main commit churn

Proposed Solution

No response

Alternatives Considered

No response

[wallstop]

Partially addressed by 66feaae (#331) — leaving this open, because the largest source of churn is untouched.

The PR body listed this issue under Closes, which was an overclaim on my part. (GitHub only applies a closing keyword to the first issue in a comma-separated list, so it did not actually close — but the claim was still wrong.)

What landed. Two of the three post-merge auto-commit flows are now one workflow. post-merge-maintenance.yml replaces update-llms-txt.yml and update-issue-template-versions.yml, running both generators in a single job that produces at most one chore: refresh generated repository metadata commit instead of up to two.

What remains — and it is the bulk of it. perf-numbers.yml still pushes its own commit:

772: git commit -m "docs(perf): auto-update dispatch throughput numbers"

That is 10 of the last 25 commits on master. Collapsing it into the same flow is harder than the other two were, and the constraints are worth writing down before anyone tries:

1. It cannot simply move into post-merge-maintenance.yml. The perf numbers come from a benchmark run on a self-hosted Windows runner holding the organization build lock and a Unity licence seat. The maintenance job is a two-minute ubuntu-latest job running two node scripts. Merging them would make every llms.txt refresh wait on — and hold — a Unity seat.
2. The loop break is load-bearing. perf-numbers.yml ignores docs/architecture/performance.md and perf-baseline.csv on push precisely because its own auto-commit uses a GitHub App token, which re-triggers workflows. Any consolidation has to preserve an equivalent break or it will benchmark itself forever.
3. The two have different cadences. Metadata regeneration is cheap and should follow every merge; a benchmark is expensive and arguably should not.

So the realistic shape is probably deferred aggregation — perf writes its numbers as an artifact, and the maintenance job commits them alongside the other generated files on the next run — rather than putting the benchmark and the generators in one job. That is a design decision, not a mechanical merge, which is why it is not bundled into a PR about CI signal integrity.