Repository navigation
CI Routing Parity 2026 09 27
This is a read-only progress measurement for #1436. It changes no workflow, runner service, admission limit, branch protection or cache policy. Windows retirement and the hosted-PR/self-hosted-main routing were already applied in #1443. The three pairs below test gate outcomes on equal Git trees; they do not establish cold-build parity, shared-host capacity or a rollout decision.
For each merged PR, the first-attempt hosted pull_request run and the
first-attempt self-hosted push run completed successfully. GitHub's run
head_sha identifies the PR source head, while the PR's trust-artifact name
identifies the synthetic merge commit actually tested. I fetched that commit
through the Git API, checked its second parent against the PR head, and
compared its Git tree with the published main commit's tree. The workflow blob
and gate-inventory blob also agree within each pair. The three workflow
blobs are all cd8c6a4946bcb0e2c2167629ae9c8b232794ef80; the inventory
was schema 5 for the first two pairs and schema 6 for the third. The required
build, roadmap and ci jobs are present in both versions.
The owner-local archive at
~/.loop-runs/transcripts/analysis/2026-09-27-m54-recovery/probes/1436-routing-parity/
contains the raw run and attempt
responses, complete attempt-job pages,
artifact listings and subject commits, six build logs, a semantic comparison
of one artifact pair, analysis scripts and SHA256SUMS. All 63 archived files
verify with sha256sum -c SHA256SUMS. summary.json has SHA-256
8fda0ac2ce700b9ff077ff0cb8db6de1faf1db2d18b1115db0273258bad73c1c.
| PR and change | Hosted PR run | Main push run | Tested tree | Main build queue, hosted / self-hosted | Build execution, hosted / self-hosted |
|---|---|---|---|---|---|
| #1606, cross-module Lean and audit | 36337863251 | 36339314603 | f8e982e140c4715f4623d805acf04e57946aab65 |
3 / 1343 s | 1393 / 887 s |
| #1609, runner admission paths | 36339562967 | 36341221828 | 96ebc376d36d1444935ab511e6e7c5da5cde67b5 |
3 / 393 s | 1579 / 888 s |
| #1615, security contract and evidence | 36343539052 | 36345139654 | 8777d3edc3eb2e68a700c489a43a23286339350b |
3 / 1 s | 1541 / 917 s |
Queue is the job's created_at to started_at; execution is started_at to
completed_at. It is not a run-created-to-run-started proxy. The self-hosted
queue ranges from one second to 22 minutes 23 seconds in these three samples;
the hosted queue is three seconds each. The small sample cannot estimate a
service-rate distribution, and a run's delayed job cannot by itself identify
CPU, memory, disk or runner-availability causation.
For every pair, build, roadmap and aggregate ci succeeded on both lanes.
Each job had the same ordered step names and conclusions across its two runs:
36/36 build steps, 6/6 roadmap steps and 3/3 aggregate steps, all success.
That includes the algebraic-geometry, stability-condition and DG axiom audits,
audit completeness, single instantiation, linters, warning and style ratchets,
source/ownership/umbrella checks, the no-sorry gate, the emitter, contract
gates, emitter reproducibility and artifact upload. The PR build jobs ran on
GitHub Actions hosted runners; the main build jobs ran on
chris-ubuntu-main-runner. The roadmap and aggregate jobs ran on hosted
runners in both lanes. This is observed outcome parity, not proof that all
internal artifacts have identical raw bytes.
I downloaded the trust artifacts for the #1606 pair for a closer comparison.
The algebraic-geometry, stability-condition and DG audit records have equal
name-and-axiom multisets (5578, 7287 and 1373 records respectively). The
first two raw audit files differ in record order; the DG file is byte-equal.
The two emission files have the same ordered 31,904 constants, modules and
counts. The declaration files have the same ordered 31,904 declarations and
counts. Their source-commit and emission-time fields differ as expected
between a synthetic PR merge and a main squash commit, and dependent hashes
therefore differ. The environment digests are equal; attest/build.json is
byte-equal. The raw comparison is pr-1606/artifact-parity.json in the archive
(SHA-256 93e797c8165295300f6bc69863753c716f4ca649cdffc465e39a1d5570862d94).
The other two pairs have gate-outcome parity but no cross-run artifact-content
comparison in this slice.
All six archived build logs say Cache restored from key, so this is a
warm-cache sample. The keys and logs are retained; a successful restore
does not prove that every source artifact was reused. Hosted and self-hosted
machines differ, and no per-job PSI, cgroup CPU/memory trace or disk-pressure
series was available for these completed runs. host-snapshot.json records
only a later point-in-time host configuration and pressure sample, not the
conditions of the six jobs. The four Ubuntu runner services share one physical
host, as the runner runbook states; four labels are not
four independent capacity pools.
At collection time, branch protection required only ci, with strict
up-to-date checking. The sampled PR build runners were hosted and the sampled
main build runners were self-hosted. The workflow has no
pull_request_target trigger. This read-only observation does not change
protection or prove a rollback. The auxiliary cache-warm workflow and dynamic
security check are separate from these three required CI jobs; their status
must not be inferred from a successful ci aggregate.
The next #1436 slice needs #1435's isolated job pickup and #1434's admission contract before a routing cutover. Capture per-physical-host queue, cgroup CPU/RAM and disk/PSI during named concurrent jobs; run a controlled cancellation/restart and a cold/warm matched revision; then rehearse the reviewed hosted-routing rollback and attach exact PR and main run identities. Until those results exist, keep the current routing and capacity limits and leave #1436 open.
Exported from docs/ci/routing-parity-2026-09-27.md at the commit that removed it from the tree.
Architecture
- Canonical roots and specialization trees
- Structural cutover ledger
- Generalization backlog
- Dependency direction
- Loop-engineering friction log
- Mathematical ownership and abstraction boundarie
- Declaration placement
- When a proposition-valued class is justified
- Enhanced spherical twists: conventions and remai
- Verification history
CI and runners
- Bounded audit batching probe, 2026-09-27
- CI1.10: audit and exception record contention, 2
- Proposed hosted-only CI cutover for CI1.07
- Mac VM runners
- Owner CI lane: self-hosted runners in a private
- CI performance baseline, 2026-09-27
- CI1.07 exact-tree Ubuntu routing sample, 2026-09
- Ubuntu workstation runners (historical topology)
Decision records
- ADR-0001 — The seam is cold: files, not runtime (superseded)
- ADR-0002 — Identity is minted in the topic repo, (superseded)
- ADR-0003 — Elaboration and axiom evidence is pro (superseded)
- ADR-0004 — The statement registry lives in the t (superseded)
- ADR-0005 — Seven axes, four values, no aggregate (superseded)
- ADR-0006 — All contract issues are filed on `bri (superseded)
- ADR-0007 — Where the contract package lives (superseded)
-
ADR-0008 —
@[cites]ships as a shared Lake dep (superseded) - ADR-0009 — The contract package lives with the e (superseded)
- ADR-0010 — How a dg category is encoded, and whe
- ADR-0011 — Hom-complexes are complexes of abelia
- ADR-0012 — One Mathlib-style source root
- ADR-0013 — Own reusable moduli infrastructure in
- ADR-0014 — Stacks, algebraic spaces and moduli:
Roadmaps