t27b-lab: judge a lane request's on-origin against a fresh fetch of origin master, not the master the lab last ran #8785

Found by the MVP acceptance lane. Root cause of lost lane receipts on t27b-lab.

Problem

lane_request() in contrib/railway/t27b-lab/lab.py builds the "on origin" set from git ls-remote --heads plus git rev-list --first-parent <master> in CLONE, where <master> is the master the lab LAST RAN. A squash commit that landed after that run is NOT_ON_ORIGIN until the lab's next master run. Its request waits as WAIT_ORIGIN and is dropped once it has waited longer than REQUEST_ORIGIN_GRACE_S. That dropped 6a7b774 (MVP v2, see #8685) on 2026-10-10. The grace went from 30 min to 2 h (see #8320 and see #8715); that softened the symptom and left the cause.

Change

• lane_request() fetches origin's REF into its remote-tracking ref (CLONE's checkout does not move) and reads the first-parent line of that fresh ref, plus the ls-remote heads. The master the lab last ran no longer matters.
• Hand-written Python stays net <= 0 lines (check_budget, see policy: hand-written code in any language but t27 may only shrink (Closes #8227) #8236).
• specs/verified/corpus_receipt.t27: the comment that says the lab judges on-origin against the master it last ran is corrected, and a test pins the verdicts for a squash commit that is newer than the last master run. request_verdict and request_exit do not change; WAIT_ORIGIN (see verified(corpus_receipt): a request the lab has not fetched yet waits 30 min instead of being dropped (Closes #8320) #8329) and the 2 h grace stay.
• Kept as they are: requests skip the fuzzer (see t27b-lab: a lane request skips the fuzzer (Closes #8267) #8268), and up to two requests run per master run (see t27b-lab: take up to two lane requests per master run (Closes #8586) #8588).

Acceptance criteria

• A scripted reproduction on t27b-lab: with CLONE at an older master M0 and a sha X on origin master's first-parent line newer than M0, master's lab.py leaves X as WAIT_ORIGIN, and drops it as NOT_ON_ORIGIN once X is older than the grace. The PR's lab.py admits X and runs it.
• A waiting request survives a lab restart. /srv/requests is on the volume, the age comes from the file's mtime, and nothing is judged before master's t27c exists.
• Both labs are redeployed right after each one publishes a run, and each lab's /status.json image.lab_py_sha equals git hash-object of master's lab.py.
• Live: a request for a just-merged squash commit, written before the lab's next master run, is admitted, not dropped.

Boundary

contrib/railway/t27b-lab/lab.py
specs/verified/corpus_receipt.t27
bootstrap/gen/rust/verified/corpus_receipt.rs
.trinity/seals/verified_VerifiedCorpusReceipt.json

User Scenarios

• Given a lab CLONE at master commit M0 and a squash commit X on origin master's first-parent line that is newer than M0, When the lab processes a lane request for X, Then the request is admitted and run (not left as WAIT_ORIGIN or dropped as NOT_ON_ORIGIN).
• Given a waiting request for commit X that is older than REQUEST_ORIGIN_GRACE_S, When the lab restarts, Then the request survives the restart and its age is still derived from the file's mtime in /srv/requests.
• Given both labs have just published a run, When they are redeployed, Then each lab's /status.json image.lab_py_sha equals git hash-object of master's lab.py.

Requirements

• FR-001: lane_request() in contrib/railway/t27b-lab/lab.py MUST fetch origin's ref into its remote-tracking ref and read the first-parent line of that fresh ref plus the ls-remote heads to judge on-origin.
• FR-002: The master the lab last ran MUST no longer be used for on-origin judgment.
• FR-003: Hand-written Python in contrib/railway/t27b-lab/lab.py MUST stay net <= 0 lines (check_budget).
• FR-004: specs/verified/corpus_receipt.t27 MUST correct the comment about judging on-origin against the master the lab last ran and MUST add a test pinning verdicts for a squash commit newer than the last master run.
• FR-005: request_verdict and request_exit MUST NOT change; WAIT_ORIGIN and the 2 h grace period MUST stay as they are.
• FR-006: Requests MUST continue to skip the fuzzer and up to two requests MUST run per master run.