hyperpolymath · hyperpolymath · May 20, 2026 · May 20, 2026
@@ -6,7 +6,7 @@
 [metadata]
 project = "burble"
 version = "1.2.0-pre"
-last-updated = "2026-05-13"
+last-updated = "2026-05-20"
 status = "active"
 
 [project-context]
@@ -44,6 +44,9 @@ doc-reality-drift = [
   "RESOLVED 2026-05-19: Formal-proof claim scoped to compile/type-check only; runtime enforcement is roadmap per ADR-0008 (Option C, PoC #55) + ADR-0007 claim discipline",
   "RESOLVED 2026-05-19: README comparison table + Status note state PTP <1us needs a PTP NIC (NTP ~1ms fallback), hardware unvalidated, per ADR-0007"
 ]
+test-isolation-62 = [
+  "RESOLVED 2026-05-20: #62 (Earn-core/#39 split-out) fully closed at the root via PRs #70/#71/#73/#76. Bucket A (port-bind races on RTSP/PTP/phc2sys) + Bucket C (Task.shutdown cross-pid ownership) cured by app-owned-GenServer unique-name pattern (#71). Bucket B (mid-run singleton death cascade) cured by Media.Peer + Coprocessor.Pipeline restart:temporary + Media.Engine try/catch :exit + Peer is_pid guard on channel_pid + HealthMesh Map.get(:port) (#76). Bucket D (SNIFBackend Logger.require + def-promotions + RoomChannel catch-all) fixed (#70). SingletonWatcher diagnostic landed in test_helper.exs (#73) — advisory, runs every suite, freezes pre-shutdown."
+]
 resolved-2026-04-16 = [
   "Opus naming/contract drift: Backend.audio_encode/4 + audio_decode/3 docstrings rewritten to state explicitly that they are PCM frame pack/unpack, NOT Opus. Added explicit Backend.opus_transcode/4 callback returning {:error, :not_implemented} on every backend (ElixirBackend, ZigBackend, SmartBackend, SNIFBackend). Added opus_available?/0 callback (always false). Pinned by opus_contract_test.exs."
 ]
@@ -161,6 +164,54 @@ open-failures = 0
 #             total non-deterministic (99..165). Cannot be
 #             authoritatively measured/validated without CI. #39 now
 #             HARD-BLOCKED on the stuck GitHub Actions queue (user).
+# 2026-05-20: #62 fully closed at the root in 4 PRs.
+#             #70 (Bucket D — genuine bugs): SNIFBackend `require
+#             Logger` + `parse_fft_result`/`prepare_ifft_input`/
+#             `parse_ifft_result` promoted defp→def (callable from
+#             tests + sibling backends); RoomChannel catch-all
+#             `join/3` clause returns {:error, %{reason:
+#             "invalid_topic"}} for non-room: topics.
+#             #71 (Bucket A — port-bind races + Bucket C — task
+#             ownership): app-owned singletons that bind ports
+#             (Burble.Timing.PTP, Burble.Transport.RTSP,
+#             Burble.Timing.Phc2sys) now honour a :name opt on
+#             start_link/1 via {name, init_opts} = Keyword.pop;
+#             tests pass `name: :"<mod>_test_<unique>"` so async
+#             test runs don't collide on registered name nor on the
+#             :gen_tcp port. llm_test.exs Task.async + Task.shutdown
+#             replaced with spawn (cross-pid ownership exit
+#             eliminated). 4 RTSP SETUP-over-TCP failures left as
+#             pre-existing parser drift (out of scope, not #62).
+#             #73 (Bucket B — diagnostic): SingletonWatcher in
+#             test_helper.exs monitors 20 app-owned singletons,
+#             records death (name+pid+reason+ms), freezes via
+#             ExUnit.after_suite before BEAM shutdown so app-
+#             teardown :DOWN cascade is not mis-recorded; advisory,
+#             does not fail CI. Caught the Bucket B cascade before
+#             the cure landed.
+#             #76 (Bucket B — cure): Burble.Media.Peer +
+#             Burble.Coprocessor.Pipeline switched to
+#             `use GenServer, restart: :temporary` (test-induced
+#             peer crashes no longer drain the supervisor's
+#             max_restarts: 3-in-5s intensity); Burble.Media.Peer
+#             guards `send(state.channel_pid, _)` with `is_pid/1`
+#             (tests pass channel_pid: nil — the :badarg was the
+#             cascade trigger); Burble.Media.Engine wraps the two
+#             DynamicSupervisor.start_child calls in try/catch :exit
+#             so a dead PeerSupervisor/CoprocessorSupervisor doesn't
+#             propagate :shutdown into the Engine; Burble.Groove.
+#             HealthMesh switched `v.port` → `Map.get(v, :port)`
+#             (peers added via report_peer_status/3 lack :port,
+#             :badkey was a secondary independent cause).
+#             Local OTP25 evidence post-PR-D: 707 tests, 134
+#             failures, NO singleton-death block emitted (every
+#             watched singleton stayed alive). 134 is inside the
+#             documented same-SHA noise band (15-38 on CI, wider
+#             locally with PR-B's test rewrites in the tree).
+#             Headline signal = absence of watcher block, not the
+#             count. OTP27 CI deferred — estate Elixir CI queue was
+#             exhausted at merge time (concurrency-pool throttle);
+#             will surface in next post-merge run on main.
 
 [crg]
 grade = "C"

@@ -10,6 +10,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Added
+- `Burble.TestSupport.SingletonWatcher` in `test/test_helper.exs` — `Process.monitor`s each of 20 app-owned singletons (PubSub, Presence, RoomRegistry/Supervisor, PeerRegistry/Supervisor, CoprocessorRegistry/Supervisor, MessageStore, NNTPSBackend, Media.Engine, Timing.{PTP,ClockCorrelator,Alignment}, Groove + HealthMesh + Feedback, Transport.RTSP, Bolt.Listener, Endpoint), reports any mid-run death (name + pid + reason + ms-since-start) to stderr at suite end, freezes via `ExUnit.after_suite/1` before BEAM shutdown so the normal app-teardown `:DOWN` cascade is not mistaken for instability. Diagnostic for #62 Bucket B; advisory (does not fail CI).
+
 ### Changed
 - README/ROADMAP claims scoped to the shipped build per ADR-0007: QUIC & SNIF marked experimental (optional NIFs disabled by default), PTP <1µs flagged hardware-gated, Idris2 proofs flagged type-check-only (runtime enforcement = ADR-0008 Option C), latency/scale flagged unbenchmarked; added a README Status section. Closes the STATE.a2ml doc-reality-drift entries (issue #51)
 
@@ -29,7 +32,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - `docs/developer/wsl-mirrored-networking.adoc` rewritten — NAT + host forwarder is the recommended WSL2 Bolt path; mirrored networking demoted to last-resort (Win11 24H2/Insider `Wsl/Service/E_UNEXPECTED` instability)
 
 ### Fixed
-- #39: Elixir test gate was dead because `Burble.Store` refused to start without VeriSimDB (first supervised child) — the whole supervision tree collapsed and every app-dependent test failed at boot. Added prod-safe, config-gated `offline_ok` (test env starts the Store in degraded offline mode); fixed 3 test-file compile errors (`snif_backend_test.exs` invalid index + nested `describe`, `room_property_test.exs` `check all` newline). Suite now compiles and boots (707 tests run); full green + gate re-arm still pending an authoritative CI run.
+- #62: test-isolation cascade fully resolved at the root (Buckets A/B/C/D). **Bucket A** (port-binding races on RTSP 19554, PTP, phc2sys ~40+ failures) and **C** (cross-pid `Task.shutdown` ownership) fixed by switching app-owned GenServers to unique-named instances per test (`PTP.start_link/1`, `RTSP.start_link/1`, `Phc2sys.start_link/1` now honour `:name` opt; tests pass `name: :"<mod>_test_<unique>"`). **Bucket B** (mid-run singleton death cascading "(EXIT) no process") root-caused via a `SingletonWatcher` GenServer in `test_helper.exs` that `Process.monitor`s each of 20 app-owned singletons and reports mid-run deaths to stderr; cured by (i) `Burble.Media.Peer` + `Burble.Coprocessor.Pipeline` flipped to `use GenServer, restart: :temporary` (so test-induced peer crashes don't drain the supervisor's `max_restarts: 3-in-5s` intensity and kill the supervisor), (ii) `Burble.Media.Peer` guarding `send(state.channel_pid, _)` with `is_pid/1` (tests pass `channel_pid: nil`, the `:badarg` was the cascade trigger), (iii) `Burble.Media.Engine` wrapping `DynamicSupervisor.start_child` in `try/catch :exit` so a dead supervisor's `:shutdown` doesn't propagate into the Engine, (iv) `Burble.Groove.HealthMesh` using `Map.get(v, :port)` instead of `v.port` (peers added via `report_peer_status/3` lack `:port`, the `:badkey` was an independent secondary cause). **Bucket D** (genuine code bugs: `SNIFBackend` missing `Logger.require` + private parser functions, `RoomChannel.join/3` non-exhaustive clauses) fixed directly. Shipped as PRs #70 (D), #71 (A+C), #73 (C diagnostic), #76 (B cure).
 - SNIF: `Burble.Coprocessor.SNIFBackend` no longer emits a compile warning for the optional `Wasmex` runtime and no longer mis-fails when it is absent — `Wasmex` is referenced via `apply/3` and `available?/0` now gates on it loadable, so kernels degrade cleanly to `ZigBackend` (mirrors the `:quicer` pattern, ADR-0004)
 
 ### Removed