fix: connecting no longer purges demo data at all (iOS parity)

[foureight84]

The bug

Seeded demo data disappears within minutes of being created, with no user action, whenever a ring is paired.

Reported as "reseed demo data seems to reset when I navigate from Today to Sleep". Navigation was a red herring — four Today↔Sleep bounces with demo counts pinned at 772/84/36 and no device event firing at all.

The cause is the reconnect cadence. Captured live while sampling the DB every 5s with the phone untouched:

08:57:53  demo=772/84/36  conn=08:52:34
09:00:52  demo=772/84/36  conn=-             <- link drops
09:00:58  demo=0/0/0      conn=09:00:54      <- reconnect wipes it

EventPersistenceSubscriber ran clearDemo() across all four tables on CONNECTED, gated only on isConnectTransition — "this came from the BLE client's own connect, not a decoder re-asserting CONNECTED". That gate correctly excludes decoder status echoes, but not a first connect from the 200th, and a paired COLMI R10 re-establishes its link roughly every five minutes.

Why it's removed outright rather than narrowed

My first cut gated the purge to first-connect only. Digging into the history showed that was still too conservative: the purge was never a deliberate product decision. It's the last surviving fragment of the original "connect = clear everything and rebuild from the ring" design, which five commits have been steadily dismantling:

commit	change
11bf6fa initial	clearDemo() on measurements/activity + clear() on all sleep (real rows)
aefa646	sleep narrowed clear() → clearDemo()
c3b669e	gate the rebuild on a real transition, not any CONNECTED
e11be6e	stop deleting stored sleep history on every connect
7649513	make "connect deletes nothing" structural

iOS — which this is a port of — has no connect-time demo purge anywhere. Its only deletion of seeded rows is user-initiated, inside SeedData. It handles the demo/real mix by detecting it (isDemo, source == "mock", DataFreshness.demo, calibrationState(isDemo:)) and adapting the UI. Demo data coexisting with a real ring is an expected state there, not something to clean up.

So this finishes the direction every prior commit was heading. Demo rows are now retired only from Settings → Privacy & Data → Clear Demo Data.

The tripwire is kept

android/AGENTS.md documents ConnectPurge as a deliberate compile-time guard against re-introducing connect-time deletion (issue #43). Deleting the type would have thrown that away, so it's kept and reduced to a single NOTHING member.

The resulting single-branch whens in EventPersistenceSubscriber and EventPersistenceIdentityTest read as dead code and are not — they're what makes re-adding a delete a compile error rather than a one-line .clear(). Both the code comment and AGENTS.md now say explicitly not to simplify them away.

AGENTS.md is updated: "a connect may retire demo rows and nothing else" → "a connect may delete nothing at all", plus a new paragraph recording the reconnect measurement and the iOS parity rationale.

Testing

Full suite: 1039 tests, 0 failures.

• no connect event may purge anything, for any family — the exhaustive-when tripwire, now over the single-member enum.
• connecting never retires demo rows, first connect or reconnect — every RingDeviceType plus the null (decoder-echo) origin.

Not yet hardware-verified. The honest test is: install, seed demo data, leave the phone alone ~10 minutes, confirm the counts survive a reconnect. Happy to run that before merge.