hertz -> doyle, todlando. PART (A) DONE at cba18cd5. No execution; builds only in my pool. Three things need your eye: your cell spec is false by design, your red-first arm points the wrong way, and the two pre-existing reds are SOLVED at zero cost -- which means I withdraw the assembly warning I gave you.

1. THE CELL: cba18cd5 on test/304-remote-friction, test-only, 55 insertions 0 deletions, trailer audited by raw grep. Compile-gate at these bytes: cargo test -p spt-daemon --lib --no-run TRUE EXIT 0, 0 errors, 0 warnings (direct capture, no pipe). Predictions for the window filed BEFORE any execution at predictions-cba18cd5.txt, Q1-Q6.

2. YOUR SPEC IS FALSE BY DESIGN, so the cell pins a different contract and says so in its own comment. You asked for Some(bound) after start_serving returns. start_serving passes broker = None, and the publish sits inside `broker.as_ref().map(|_| { BOUND_DOCS_PORT.store(bound, ...); BoundDocsPort })`, so a brokerless listener NEVER stores and bound_docs_port() stays None however successfully it bound. The module comment above BOUND_DOCS_PORT states that as the intent: "Only the broker-backed listener publishes here; standalone fixture listeners must not advertise their ports as the daemon's serving surface." Nothing tested it. So the cell asserts THE GATE -- a real ephemeral bind, then None -- which is the claim with a live failure mode: hoist that store out of the map and every temp-dir fixture listener in this suite advertises itself as the daemon.
   Your third assertion is also unreachable as specified: BoundDocsPort is held in the serving task for the lifetime of accept_loop, so a unit cell cannot drop it without a stop handle and there is none to take. I assert the retirement ON THE GUARD instead (store, Some, drop, None -- which also covers the 0-sentinel mapping in BOTH directions), and the comment labels it as the reachable half of that claim rather than passing it off as a server test.
   Your two near cells are as you described: port_resolution_env_config_default is config precedence with no daemon, and rig_ephemeral_docs_port_retires_a_configured_collision calls start_serving but never reads bound_docs_port. Neither asserts the claim, so there is no duplication.

3. YOUR RED-FIRST ARM POINTS THE WRONG WAY, and it would have produced a green arm read as a control. "The store commented out" removes the publish -- but my gate assert EXPECTS None, so deleting the store leaves it PASSING. The mutation that falsifies the gate has to make a brokerless listener PUBLISH: hoist the store OUT of the broker map. Filed as ARM R-GATE, predicted RED at the gate assert with left Some(<the ephemeral port start_serving returned>), right None. The retirement half has its own arm (ARM R-DROP: delete the store in impl Drop) predicted RED at the final assert. Gate arm first if only one build is affordable, because it is the claim nothing else in the tree covers.

4. THE TWO PRE-EXISTING REDS ARE SOLVED, no window, no build, no test -- and it is MY RIG, not the product. Your TMP candidate was right and it is now measured with a control:
     git -C <my producer's TMP: EVID/temp/w5-lib-full> rev-parse --show-toplevel
       -> C:/Users/decid/Documents/projects/spt-core , remote https://github.com/BigscreenVR/spt-bs-core.git
     git -C C:/Users/decid/AppData/Local/Temp rev-parse --show-toplevel
       -> fatal: not a git repository        (todlando's placement, and CI's)
   lane_run.py set TMP/TEMP/TMPDIR/RUNNER_TEMP to EVID/temp/<label>, which lives INSIDE the spt-core checkout. So every tempfile::tempdir() fixture in my producers sat under a git work tree, git discovery walked up, found the repo, and project_id_for_dir -- REMOTE-URL FIRST -- handed back this repo's identity to cells whose fixture requires a directory outside any repo.
   todlando's string derivation closes it and I re-checked it rather than relaying: slug(remote) == "github-com-bigscreenvr-spt-bs-core" (registryhost's exact observed value) and display_from_url(remote) == "spt-bs-core" (projwriter's exact observed value). The folder-name fallback could not have produced either -- the candidate folder names here are 304-product, hertz-304 and spt-core, and none appears. So git got as far as step ONE of the derivation; the fallback path is excluded, not merely disfavoured.
   I WITHDRAW THE ASSEMBLY WARNING. I told you the golden lib lane would red on these two at the assembled tree. That was wrong and todlando's correction was right: the cause is rig-local TMP placement, the golden runner has its own TMP, and his own unfiltered run on this box reds on neither. The honest line for deployah is "two cells whose correctness depends on an ambient git context; verdict environment-shaped; measured PASS in two unfiltered runs with a default TMP and FAIL in one with an in-repo TMP, on identical bytes" -- not a prediction about the golden lane. Your ci.yml:143 datum is what killed my first mechanism, and it killed it properly: in-repo CWD was never the axis, the fixture's own location was.
   RIG FIXED, declared: the base now resolves outside the tree (AppData\Local\Temp\hertz-304-rig\<label>), per-label isolation and reapability unchanged, pre-change file kept as lane_run.py.bak-tmpfix, and HERTZ_RIG_TMP_IN_REPO=1 restores the old placement ON PURPOSE as the negative control. That gives your requested (a)/(b) arms directly: ARM A default predicts both cells PASS, ARM B predicts both FAIL at the exact lines and strings above. If ARM A fails the mechanism is incomplete; if ARM B passes it is wrong and I withdraw it too.

5. Q4 predicts the WHOLE-SUITE number this time: 988 run, 988 passed, 0 failed (987 + my cell, minus the two TMP artifacts), with the cfg-gap assumption named as the limit. Last window I predicted a total from one module's delta and was short by six; that is the specific mistake I am not repeating.

6. NOTED FOR HONESTY, not as an alarm: the compile-gate log for this sha carries "Blocking waiting for file lock on build directory". Your workspace check is in a fresh pool, so it is not you; it is the editor's rust-analyzer building into this lane's pool on its own schedule (IR-100), and it was name-keyed present in my admission. An unexplained lock wait is exactly what a foreign builder looks like, so it gets named rather than skipped.

EXECUTION WAITS FOR YOUR WORD after your workspace check on d0a74461 ends. On the word my window is: the new cell by name, ARM R-GATE (and R-DROP if affordable) with the mutation reverted and verified by oid between arms, then the two pre-existing cells under ARM A and ARM B. START will ride the launcher as before.