tenant.openagents/omega

Give the issue-31 host pump a Full Auto reading it measured (omega#97)

Give the issue-31 host pump a Full Auto reading it measured (omega#97) omega#49 and omega#91 both closed carrying the same named substitution: "no `omega-effectd` daemon is attached to this process, so the Full Auto reading is supplied here rather than polled". That substitution is the whole of omega#97 on the host side, and the obvious way to fill the mobile rows — replaying `fixtures/live-omega-effectd.get_run.json` — would make a recorded fixture the host authority for a device proof, which omega#49's exit forbids in as many words. So the reading is measured instead. `issue31_observation` reads a real running daemon through the five read methods (`list_runs`, `get_run`, `get_capacity`, `get_report`, `get_receipt`). It never calls `start`, `pause`, `resume`, `retry`, or `stop`: reading a run registry is not the same act as directing one, and Full Auto authority does not begin on a path a model can reach. Three properties, following omega#91: - **Host-produced means no input path.** `generated_at_ms` is now a private field stamped from one `now()` reading inside `Issue31FullAutoReading::observed`, which has no parameter for it. A fixture-authored reading is not forbidden by a comment, it is unexpressible from any production path — `at_recorded_instant` is `#[cfg(test)]` and does not exist in a release binary. `Default` is dropped for the same reason: a default reading is one nobody measured, stamped 1970. - **Refused rather than shortened.** A run the daemon listed and then would not describe refuses the whole reading. Publishing the shorter list would drop a run the host is holding, and a run that vanishes from a snapshot reads on the phone as a run that ended. - **One reading path.** The polling was inlined in `panel.rs`, reachable only from a GPUI view. `panel.rs` now calls the same observer, so the desktop and the phone cannot look at two different readings of one daemon. A host that could not look still publishes silence, not an empty view; a host that looked and found nothing still publishes zero runs. Each failure carries its own token so the two are never collapsed. Proof: `a_running_daemon_supplies_the_reading_a_paired_device_reads_on_a_live_relay` spawns the packaged `omega-effectd` under a data root of its own, reads it, and publishes the omega#47 snapshot and detail to a paired device on the DEPLOYED `wss://relay.openagents.com`. Measured 7 lanes, 0 runs, 0 accounts; the relay drained the outbox. `MockRelayAdapter` is not used. Ten falsifications, each probed on disk and confirmed rebuilt before its test ran; all ten went red for the named assertion. The first attempt at one of them wrote before its probe raised and contaminated the next three — caught, reverted, and re-run from a pinned baseline. The fourteen byte-shared `openagents.omega.issue31.*` fixtures are untouched. No new `OMEGA-DELTA`: this is new Omega surface, not a changed inherited default. This is the host half. The device half is still open on omega#97: `readIssue31FullAutoProjection` is called exactly once in the mobile app, with `(null, null)`, and `home-screen.tsx` never passes `hostAdjunct`, so the phone still opens none of what the host now publishes. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>