bitchat/bitchat.xcodeproj
jack cdebdd9347
Link layer slice 4: deterministic multi-node mesh simulation (and the panic-announce bug it caught) (#1548)
* Cohere per-link Noise auth and rebind containment into BLELinkAuthState

The authenticated-link owners, the reconnect revalidation policy, and
the two rebind-containment cooldowns were four loose bleQueue-owned
maps whose invariants lived in call-site discipline: every teardown
path had to remember to retire the proof AND close the revalidation
epoch (the pair appeared seven times), and both cooldowns hand-rolled
the same prune-check-record dance. BLELinkAuthState owns them as whole
transitions — retireLink, retireLinks(ownedBy:), permitRebind,
permitRedundantRetirement — with the ownership question (bleQueue
today, engine after the option-B flip) answered in one place.

No behavior change; the one call-site reordering (redundant retirement
computes the survivor before the cooldown check instead of after) is
outcome-equivalent since the cooldown only ever recorded when a
survivor existed.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* Split identity-link bindings out of the physical link store

BLELinkStateStore owned two different kinds of truth: what physical
links exist (CB handles, connect lifecycles, characteristics, stream
assemblers) and who each link belongs to (peer bindings in both roles
plus the preferred-peripheral reverse map for directed sends and fanout
collapse). The bindings now live on BLELinkBindings — same bleQueue
ownership, whole-transition methods, direct tests for the rotation
reverse-map cleanup and the preferred-link survivor repair that were
previously only exercised end to end. Composed operations that need
both truths (remove-with-repair, direct link state, the subscribed-
central snapshot, bind-only-live-links) live on the transport as
explicitly bleQueue-confined helpers.

This is the structural half of the option-B boundary flip
(docs/BLE-ARCHITECTURE-V3.md): ownership of the bindings can now move
to the engine without touching what-links-exist. An audit of every
physical clear/remove found three sites (emergency clear, both
unauthorized branches) that needed explicit binding-clear pairing under
the split — each now clears both.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* Fix iOS-gated constructors and preserve containment cooldowns on reset

CI caught what the macOS SwiftPM build cannot see: two #if os(iOS)
sites still passed the peerID field that slice B1 removed from
BLEPeripheralLinkState (willRestoreState in BLEService and
armPendingBackgroundConnects in BLERadioController). Both fixed and
verified with a local iOS simulator xcodebuild.

Codex also caught a real regression: BLELinkAuthState.removeAll()
cleared the rebind/retirement cooldown maps, which the original panic
and emergency reset paths deliberately left alive. A stable
CoreBluetooth UUID must not earn a fresh rebind allowance just because
the session state around it was wiped. removeAll() now clears only the
proofs and revalidation epochs, and BLELinkAuthStateTests pins the
survival invariant along with the other auth-state transitions.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* Link layer slice 3: the option-B domain flip — bindings and link-auth move to the engine

The identity domain (BLELinkBindings + BLELinkAuthState) is now owned
by the engine queue, with a DEBUG dispatchPrecondition trapping any
access from another queue. bleQueue keeps only physical link state.

What changed shape:

- Receive path is sans-I/O: bleQueue decodes frames and hands
  (packet, linkID) up through ingestDecodedPacket (panic lifecycle
  captured at the handoff); attributeAndHandlePacket resolves the
  sender binding, rejects spoofed senders, applies raw-announce
  binding, and records ingress on the engine. Per-link frame order is
  preserved end to end (both queues serial), which supersedes the old
  batch-local TOCTOU binding in the notification path.
- The rotation rebind is one engine slot: containment checks, proof
  retirement, binding flip, reconnect decision, and rotated-identity
  retirement run straight-line; only CoreBluetooth cancels hop to
  bleQueue. The engine->bleQueue->engine ping-pong is gone, along with
  the _test_afterVerifiedDirectRebindEnqueued pause hook — the test
  that used it now asserts the atomicity directly (a paused engine
  wedged the old gate design into a three-queue deadlock).
- Authenticated-send eligibility (notifyOrEnqueueIfAccepted,
  writeOrEnqueueIfAccepted) is checked on the engine, serialized
  against rebinds by construction; only physical admission
  (updateValue/write/backpressure) runs on bleQueue.
- Teardown splits into discardPeripheralLinkPhysical (bleQueue, inline
  in the delegates) + retirePeripheralLinkIdentity (engine hop with
  survivor repair reading liveness via readLinkState). A binding can
  briefly outlive its physical link; liveness queries join against the
  physical store and the queued retirement converges the two.
- Gossip delegate sends enter the engine via onEngine — safe because
  mesh.sync sits above the engine in the sync order (production engine
  code only async-dispatches into the manager).
- checkPeerConnectivity rides an engine slot from the bleQueue
  maintenance tick.

No wire changes. 1,974 tests green (parallel and serial), iOS
simulator build clean, Periphery clean.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* Link layer slice 4: deterministic multi-node mesh simulation — and the panic-announce bug it caught

SimulatedMesh wires real CoreBluetooth-free BLEService engines
edge-to-edge through the outbound packet tap and _test_ingestFrame
(the production attribution path the B2 flip created), with per-edge
synthetic link IDs and manual-scheduler time. Five multi-node tests
run in ~40ms with no wall-clock waits:

- announce exchange binds simulated links and connects peers
- Noise sessions establish end-to-end (real crypto, both directions)
- a public message relays across a line topology inside a TTL/frame
  budget (storm bound asserted)
- an 8x duplicate flood delivers exactly once
- a panic rotation rebinds the survivor's link exactly once and stays
  — the scenario that previously needed two phones and log archaeology

Fidelity boundary (documented in the harness): no physical links, so
fanout planning and backpressure are not exercised; attribution,
binding, dedup, TTL, relay decisions, and sessions are the real
engine code.

The simulator found a real bug on its first run: the forced-announce
throttle's lastSent survived a panic, so a rotation within
bleForceAnnounceMinIntervalSeconds of the last announce silently
swallowed the new identity's announce — leaving it invisible to the
mesh until the next maintenance cycle. Today's device test only
passed because the previous announce happened to be minutes old.
BLEAnnounceThrottle gains reset(), called from the panic slot so the
rotated identity owes no throttle debt; pinned by a unit test and the
mesh rotation test.

New DEBUG seams: _test_ingestFrame (production ingress attribution),
_test_forceAnnounce, _test_fenceEngine.

1,980 tests green, Periphery clean, iOS simulator build clean.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

---------

Co-authored-by: jack <jackjackbits@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-07-30 01:14:19 +01:00
..