From 5f5c34fb9cf68c9a4a8e60b6c02dae7d5fa81e1c Mon Sep 17 00:00:00 2001 From: Luke Curley Date: Wed, 4 Nov 2026 22:21:24 -0800 Subject: [PATCH 1/2] quest: plan follow-ups from the 2026-11-05 merge batch Co-Authored-By: Claude Opus 5.5 --- quest/m0/ci/README.md | 2 ++ quest/m0/ci/macos-kio.md | 16 ++++++++++ quest/m0/ci/nix-builder-userns.md | 25 ++++++++++++++++ quest/m0/test-flakes-2/README.md | 2 ++ .../test-flakes-2/ffi-datagram-roundtrip.md | 21 +++++++++++++ .../m0/test-flakes-2/release-test-offline.md | 18 +++++++++++ quest/m1/README.md | 2 ++ quest/m1/bbr-collapse-estimate.md | 25 ++++++++++++++++ quest/m1/route-swap-keeps-incumbent.md | 30 +++++++++++++++++++ quest/m2/README.md | 1 + quest/m2/resume-redesign.md | 9 ++++++ quest/m2/watch-symlink-targets.md | 18 +++++++++++ 12 files changed, 169 insertions(+) create mode 100644 quest/m0/ci/macos-kio.md create mode 100644 quest/m0/ci/nix-builder-userns.md create mode 100644 quest/m0/test-flakes-2/ffi-datagram-roundtrip.md create mode 100644 quest/m0/test-flakes-2/release-test-offline.md create mode 100644 quest/m1/bbr-collapse-estimate.md create mode 100644 quest/m1/route-swap-keeps-incumbent.md create mode 100644 quest/m2/watch-symlink-targets.md diff --git a/quest/m0/ci/README.md b/quest/m0/ci/README.md index 6b84f40ebd1..3ab6cf8197e 100644 --- a/quest/m0/ci/README.md +++ b/quest/m0/ci/README.md @@ -186,3 +186,5 @@ The historical samples locate candidates; they are not controlled benchmarks. - [Measure the Linux cache effect of native media exclusions](/quest/m0/ci/hakari-followups.md) - scheduled Linux jobs measure cache reuse after unrelated tests stop compiling PipeWire and libclang - [A release merge opens no stale release PR](/quest/m0/ci/release-pr-after-release.md) - `just rs release` publishes before it proposes, so a merged release isn't re-proposed - [Build two queue entries at once if it helps](/quest/m0/ci/queue-concurrency.md) - a measured trial of `maximumEntriesToBuild: 2` against the serial lane's burst latency +- [macOS CI runs kio's tests](/quest/m0/ci/macos-kio.md) - `just rs macos-test` includes kio +- [The CI host's Nix builders may map user namespaces](/quest/m0/ci/nix-builder-userns.md) - `check flake` runs its user-namespace tests on the host instead of skipping diff --git a/quest/m0/ci/macos-kio.md b/quest/m0/ci/macos-kio.md new file mode 100644 index 00000000000..c947390a165 --- /dev/null +++ b/quest/m0/ci/macos-kio.md @@ -0,0 +1,16 @@ +# [XS] macOS CI runs kio's tests + +## Goal + +The macOS test job runs `kio`'s tests, so platform-dependent kio failures, +such as the `Mutex` size assumption +[moq-dev/moq#7914](https://github.com/moq-dev/moq/pull/7914) fixed, fail CI +instead of reaching main. + +## Plan + +Decided 2026-11-05 (maintainer, follow-up from #7914): ✅ m0. + +`just rs macos-test` (`rs/justfile`) tests hang, moq-tokio, moq-sock, +moq-video, and moq-auth, but not kio. Add it, and check the job's time stays +reasonable. diff --git a/quest/m0/ci/nix-builder-userns.md b/quest/m0/ci/nix-builder-userns.md new file mode 100644 index 00000000000..c15e3f27dfa --- /dev/null +++ b/quest/m0/ci/nix-builder-userns.md @@ -0,0 +1,25 @@ +# [S] The CI host's Nix builders may map user namespaces + +## Goal + +On the maintainer's CI host, `check flake` runs its user-namespace checks +instead of skipping, so `just ci` there covers the runner VM image tests. + +## Plan + +Decided 2026-11-05 (maintainer, follow-up seen in +[moq-dev/moq#7918](https://github.com/moq-dev/moq/pull/7918), +[moq-dev/moq#7684](https://github.com/moq-dev/moq/pull/7684), +[moq-dev/moq#7858](https://github.com/moq-dev/moq/pull/7858), and +[moq-dev/moq#7883](https://github.com/moq-dev/moq/pull/7883)): ✅ m0. + +Builders there refuse `unshare --map-root-user` with +`unshare: write failed /proc/self/uid_map: Operation not permitted`. +[moq-dev/moq#7763](https://github.com/moq-dev/moq/pull/7763) retired the +earlier quest because the local check already detects this and skips +(`nix/runner/userns-probe.nix`, `sh/dispatch.sh`). This quest fixes the host +instead, so nothing is skipped. The host is configured in the kixelated/nixos +repository and reachable as `ssh nixos`. Change the Nix builder sandbox +settings there, apply when the merge queue is idle, and check that +`nix/runner/userns-probe.nix` succeeds and `check flake` runs its user-namespace +tests. diff --git a/quest/m0/test-flakes-2/README.md b/quest/m0/test-flakes-2/README.md index 76f225a7a5d..6c21678a2fd 100644 --- a/quest/m0/test-flakes-2/README.md +++ b/quest/m0/test-flakes-2/README.md @@ -100,3 +100,5 @@ Public API: none. Wire: none. - [moq-gst's interleaved restart test passes under load](/quest/m0/test-flakes-2/gst-reconnect-interleave.md) - `restarts_under_reconnects_interleave_on_one_pad` passes in loaded runs and the merge queue, its missed delivery and its stray `Eos` each fixed at the cause - [reload.test.ts's close-drain tests pass under load](/quest/m0/test-flakes-2/reload-close-drain.md) - the drain cases in `js/net`'s reload tests pass in loaded multi-file runs - [dispatch-query's terminate test survives load](/quest/m0/test-flakes-2/dispatch-query-terminate.md) - `test_terminating_the_run_stops_the_running_phase` passes reliably under host load, re-enabled once fixed +- [release.test.ts never reaches the network](/quest/m0/test-flakes-2/release-test-offline.md) - the JSR release test fakes its registry lookup and passes offline +- [moq-ffi's datagram tests don't hinge on one datagram](/quest/m0/test-flakes-2/ffi-datagram-roundtrip.md) - the default-connection datagram tests pass under CI load diff --git a/quest/m0/test-flakes-2/ffi-datagram-roundtrip.md b/quest/m0/test-flakes-2/ffi-datagram-roundtrip.md new file mode 100644 index 00000000000..36f4ca9c5a8 --- /dev/null +++ b/quest/m0/test-flakes-2/ffi-datagram-roundtrip.md @@ -0,0 +1,21 @@ +# [S] moq-ffi's datagram tests don't hinge on one datagram + +## Goal + +`datagrams_cross_a_default_connection` and its `requested_` sibling +(`rs/moq-ffi/src/test.rs`) pass reliably under CI load, without depending on +a single unreliable datagram arriving. + +## Plan + +Decided 2026-11-05 (maintainer, follow-up from +[moq-dev/moq#7914](https://github.com/moq-dev/moq/pull/7914)): ✅ m0. + +The test sends one datagram over a real connection and waits 10 s for it; +it timed out in Check on #7914 while passing 30 of 30 runs on an idle host. +Confirm the cause first. If it's loss, keep writing datagrams until the first +arrives and assert on what arrives, rather than waiting longer. If it's the +datagram answer race planned in +[moq-dev/moq#7929](https://github.com/moq-dev/moq/pull/7929) +(`quest/m1/datagram-tracks/answer-race.md`), say so here and wait on that +quest instead. diff --git a/quest/m0/test-flakes-2/release-test-offline.md b/quest/m0/test-flakes-2/release-test-offline.md new file mode 100644 index 00000000000..c2a08f3e5dd --- /dev/null +++ b/quest/m0/test-flakes-2/release-test-offline.md @@ -0,0 +1,18 @@ +# [XS] release.test.ts never reaches the network + +## Goal + +`js/common/release.test.ts` runs without network access and never times out +on a live registry lookup, so it can't drop an unrelated PR from the merge +queue. + +## Plan + +Decided 2026-11-05 (maintainer, follow-up from the queue drop of +[moq-dev/moq#7867](https://github.com/moq-dev/moq/pull/7867)): ✅ m0. + +`the JSR token reaches only deno publish` timed out at 5 s in a +`merge_group` build: its fixture runs `release.ts` with `jsr=true` and +`dryRun=false`, which makes a live jsr.io metadata lookup. The land fixer +reproduced the timeout under `unshare -rn`. Inject that lookup, the way the +test already fakes the publish commands, so the whole suite passes offline. diff --git a/quest/m1/README.md b/quest/m1/README.md index 23dc7549706..477f349cf4f 100644 --- a/quest/m1/README.md +++ b/quest/m1/README.md @@ -159,6 +159,7 @@ their end report, so an embedder's drain alert pages on real cuts. - [Sockets](/quest/m1/socket/README.md) - a generic dial/accept `Socket` in moq-net, Rust, JS, and Python, with tracks as streams; moq-boy and CARLA move onto it - [A smaller, safer API than release](/quest/m1/api-audit/README.md) - every surface added since release has a consumer or is gone, overlapping operations are collapsed, names mirror across Rust, JS, and the bindings, and every break says how to migrate - [Native rung disable proof](/quest/m1/simulcast-rung-disable.md) - a shaped uplink through moq-ffi proves the top rendition stops and resumes over real BBR3 probing +- [A collapsing uplink lowers the bandwidth estimate promptly](/quest/m1/bbr-collapse-estimate.md) - the reported estimate honors BBR3's loss bound, so rungs disable within BBR3's own response time - [Audio jitter target](/quest/m1/audio-jitter-target/README.md) - the playout target is default in both languages; time-stretched browser playback without steady-state underruns and a native trace replay remain - [A/V clock](/quest/m1/av-clock.md) - the audio playhead drives Sync.reference while audio plays, through per-track sync handles - [Container reads drop backfill](/quest/m1/read-backfill.md) - the JS container consumer drops groups below its read cursor like Rust @@ -179,6 +180,7 @@ their end report, so an embedder's drain alert pages on real cuts. - [Poll-first async](/quest/m1/poll-first/README.md) - library crates compile no coroutines by default: async wrappers are generated from `poll_*` and gated behind an `async` feature - [Ordered reads can skip owed holes, and snapshot latest() does](/quest/m1/ordered-owed-holes.md) - moq-net's `Ordered` gains a read that skips past an owed hole to the newest buffered group, while keeping the reader's bounds - [The relay reports native sessions it force-closes at the drain deadline](/quest/m1/relay-drain-cuts.md) - the drain's public outcome and live tally separate sessions cut while open from sessions only finishing their end report, so an embedder's drain alert pages on real cuts +- [A replacement that fails before delivering never ends a live track](/quest/m1/route-swap-keeps-incumbent.md) - a route change whose replacement aborts before its first group leaves readers on the source that was serving - [A Rust relay never replays from group 0 to a lite-03/04 subscriber](/quest/m1/relay-inflight-replay.md) - A Rust relay or publisher serving a lite-03 or lite-04 SUBSCRIBE from a track copy that is still resolving starts at the newest group once the copy resolves - [Rename moq-rtc to moq-webrtc](/quest/m1/webrtc-rename.md) - one published `moq-webrtc` crate holds the gateway, the ICE driver, and the data-channel transport - [Vendor and trim the WebRTC stack](/quest/m1/webrtc-vendor.md) - str0m, dimpl, `is`, and sctp-proto live in moq-webrtc with unused code removed, so moq-rtc can take rind diff --git a/quest/m1/bbr-collapse-estimate.md b/quest/m1/bbr-collapse-estimate.md new file mode 100644 index 00000000000..676ea5aa7ea --- /dev/null +++ b/quest/m1/bbr-collapse-estimate.md @@ -0,0 +1,25 @@ +# [M] A collapsing uplink lowers the bandwidth estimate promptly + +## Goal + +When a path's capacity collapses, the bandwidth estimate a sender reports +(the grant simulcast rungs follow) falls as soon as BBR3's own loss response +cuts, instead of holding the old rate for seconds. A before-and-after +measurement with the sparse-sender shaper harness shows the drop. + +## Plan + +Decided 2026-11-05 (maintainer, follow-up from +[moq-dev/moq#7671](https://github.com/moq-dev/moq/pull/7671)): ✅ the +reported estimate also honors BBR3's loss-driven bound; m1, beside the native +rung-disable proof. Not chosen: m2, or an investigation-only quest. + +Evidence from #7671's sparse harness: after the path fell from 2 Mbit/s to +100 kbit/s, the estimate held 1.32 Mbit/s for about 11 s, and in the FFI run +the top rung disabled 5.6 s after the 500 kbit/s step. The probe idles while +the estimate covers the target, and the reported estimate +(`max_bw.min(bw_delay)`) ignores the loss bound BBR3 keeps +(`inflight_longterm` and its short-term bounds in +`rs/moq-congestion/src/bbr3/mod.rs`). Measure first, change second, and +record both runs. Keep the change inside the reported estimate unless the +measurement shows the control loop itself is late. diff --git a/quest/m1/route-swap-keeps-incumbent.md b/quest/m1/route-swap-keeps-incumbent.md new file mode 100644 index 00000000000..cd6ea379a30 --- /dev/null +++ b/quest/m1/route-swap-keeps-incumbent.md @@ -0,0 +1,30 @@ +# [M] A replacement that fails before delivering never ends a live track + +## Goal + +When a route change splices a replacement source under a front, and the +replacement answers with metadata and then aborts before delivering any +group, readers keep receiving from the source that was serving before. The +track ends only when no source is left to serve it. + +Not in scope: the producer-replacement redesign +([Resume as producer replacement](/quest/m2/resume-redesign.md)), which +replaces this machinery later. + +## Plan + +Decided 2026-11-05 (maintainer, follow-up from +[moq-dev/moq#7789](https://github.com/moq-dev/moq/pull/7789)): ✅ its own m1 +quest, fixed now. Not chosen: leaving it to the m2 redesign. + +#7789 fixed the case where the replacement aborts before the front reads its +answer, which now counts as a refusal. The remaining race is an abort after +the front's poll. `Front::track_info` (`rs/moq-net/src/model/front.rs`) +clears `draining` on splice, so the later `track_ended` refusal has no +incumbent to return to, and the track ends with the replacement's error. +Main has it too. Recommended fix: keep the incumbent's copy until the +replacement delivers its first group, then release it. Reproduce it first +with a test beside #7789's +`a_replacement_aborted_without_history_keeps_the_incumbent` +(`rs/moq-net/src/model/origin.rs`), and check whether `js/net` has the same +race. diff --git a/quest/m2/README.md b/quest/m2/README.md index 554e44ef40f..6e436126d6c 100644 --- a/quest/m2/README.md +++ b/quest/m2/README.md @@ -304,3 +304,4 @@ lite-07; bitrate enforcement is not part of this release. - [The publish support panel lists only codecs publishing accepts](/quest/m2/publish-support-probe.md) - the settings panel runs the publisher's Annex B encoder probe - [A hung Python test fails instead of hanging the suite](/quest/m2/py-test-timeouts.md) - one pytest-timeout bound covers every `py/` test - [Chromium's Opus encoder stamps chunks from their input](/quest/m2/chromium-opus-timestamps.md) - waiting on Chromium; unlocks native `usedtx` for push-to-talk +- [TLS reload follows symlinked certificate files](/quest/m2/watch-symlink-targets.md) - in-place rewrites of symlink targets in other directories reload roots, served chains, and client certificates diff --git a/quest/m2/resume-redesign.md b/quest/m2/resume-redesign.md index d7ebbd56226..0966f8fa422 100644 --- a/quest/m2/resume-redesign.md +++ b/quest/m2/resume-redesign.md @@ -79,6 +79,15 @@ measure the change. Public API: none. Wire: none. +Decided 2026-11-05 (maintainer, follow-up from +[moq-dev/moq#7736](https://github.com/moq-dev/moq/pull/7736)): ✅ fold this +case into the redesign instead of patching code it deletes. In +`js/net/src/resume.ts`, `#update` skips a copy whose upstream subscription +closed, so a closed copy that still holds groups keeps picking under the +reader's old `delay.max` after an update. It was found by review and not +reproduced. The redesign's tests must cover a preference update after the +serving copy's upstream closed, followed by a read. + ## Required - [Flights](/quest/m1/unbounded-groups/flights.md) - the paced copy and handover `Recover` folds into diff --git a/quest/m2/watch-symlink-targets.md b/quest/m2/watch-symlink-targets.md new file mode 100644 index 00000000000..8e5a6150dc9 --- /dev/null +++ b/quest/m2/watch-symlink-targets.md @@ -0,0 +1,18 @@ +# [S] TLS reload follows symlinked certificate files + +## Goal + +Rewriting a certificate or key in place reloads it even when the configured +path is a symlink whose target lives in another directory, as Kubernetes +secret mounts do. This covers roots, served chains, and client certificates. + +## Plan + +Decided 2026-11-05 (maintainer, follow-up from +[moq-dev/moq#7906](https://github.com/moq-dev/moq/pull/7906)): ✅ m2. + +`watch::callback` (`rs/moq-tokio/src/watch.rs`) watches each configured +file's directory, so a rewrite of a symlink target elsewhere goes unseen. +Also watch the directories the configured paths resolve to, and re-resolve +them when a link changes, so a swapped `..data` link is followed. Tests drive +the watcher with controlled events, not wall-clock waits. From d9f9c13c62480e3d0307958817e564b7e74e83ad Mon Sep 17 00:00:00 2001 From: Luke Curley Date: Wed, 4 Nov 2026 22:26:46 -0800 Subject: [PATCH 2/2] quest: correct the userns host, the symlink watch gap, and the BBR bounds nix-builder-userns targeted the NixOS host, whose builders already pass the probe; the refusal is AppArmor's userns restriction on the Ubuntu desktop host, and `just ci` fails rather than skips there. watch-symlink-targets claimed main misses a Kubernetes `..data` swap, which its filter already catches; the gap is an in-place write to a target outside the watched directory. bbr-collapse-estimate now names `bw_shortterm` as the rate bound the reported estimate omits. Co-Authored-By: Claude Opus 5.5 --- quest/m0/ci/README.md | 2 +- quest/m0/ci/nix-builder-userns.md | 31 ++++++++++++++++++------------- quest/m1/bbr-collapse-estimate.md | 11 ++++++----- quest/m2/watch-symlink-targets.md | 16 +++++++++------- 4 files changed, 34 insertions(+), 26 deletions(-) diff --git a/quest/m0/ci/README.md b/quest/m0/ci/README.md index 3ab6cf8197e..94219ba2d2b 100644 --- a/quest/m0/ci/README.md +++ b/quest/m0/ci/README.md @@ -187,4 +187,4 @@ The historical samples locate candidates; they are not controlled benchmarks. - [A release merge opens no stale release PR](/quest/m0/ci/release-pr-after-release.md) - `just rs release` publishes before it proposes, so a merged release isn't re-proposed - [Build two queue entries at once if it helps](/quest/m0/ci/queue-concurrency.md) - a measured trial of `maximumEntriesToBuild: 2` against the serial lane's burst latency - [macOS CI runs kio's tests](/quest/m0/ci/macos-kio.md) - `just rs macos-test` includes kio -- [The CI host's Nix builders may map user namespaces](/quest/m0/ci/nix-builder-userns.md) - `check flake` runs its user-namespace tests on the host instead of skipping +- [The desktop host's Nix builders may map user namespaces](/quest/m0/ci/nix-builder-userns.md) - agents' local `just ci` builds the runner image instead of failing `check flake` diff --git a/quest/m0/ci/nix-builder-userns.md b/quest/m0/ci/nix-builder-userns.md index c15e3f27dfa..ffb22d075a6 100644 --- a/quest/m0/ci/nix-builder-userns.md +++ b/quest/m0/ci/nix-builder-userns.md @@ -1,9 +1,10 @@ -# [S] The CI host's Nix builders may map user namespaces +# [S] The desktop host's Nix builders may map user namespaces ## Goal -On the maintainer's CI host, `check flake` runs its user-namespace checks -instead of skipping, so `just ci` there covers the runner VM image tests. +On the desktop host, where agents run their local gates, `just ci` builds the +flake's runner image instead of failing `check flake`, and `just check` runs +it instead of skipping. ## Plan @@ -13,13 +14,17 @@ Decided 2026-11-05 (maintainer, follow-up seen in [moq-dev/moq#7858](https://github.com/moq-dev/moq/pull/7858), and [moq-dev/moq#7883](https://github.com/moq-dev/moq/pull/7883)): ✅ m0. -Builders there refuse `unshare --map-root-user` with -`unshare: write failed /proc/self/uid_map: Operation not permitted`. -[moq-dev/moq#7763](https://github.com/moq-dev/moq/pull/7763) retired the -earlier quest because the local check already detects this and skips -(`nix/runner/userns-probe.nix`, `sh/dispatch.sh`). This quest fixes the host -instead, so nothing is skipped. The host is configured in the kixelated/nixos -repository and reachable as `ssh nixos`. Change the Nix builder sandbox -settings there, apply when the merge queue is idle, and check that -`nix/runner/userns-probe.nix` succeeds and `check flake` runs its user-namespace -tests. +Each of those gates failed `check flake` while building `github-runner`, with +`unshare: write failed /proc/self/uid_map: Operation not permitted`. `just ci` +sets `CI_STRICT`, so the check fails there; without it, +`nix/runner/userns-probe.nix` and `sh/dispatch.sh` detect the refusal and skip, +which is why [moq-dev/moq#7763](https://github.com/moq-dev/moq/pull/7763) +retired the earlier quest. This quest fixes the host instead, so nothing fails +or skips. + +As of 2026-11-05 the probe fails on the desktop host (Ubuntu 26.04, where +`kernel.apparmor_restrict_unprivileged_userns = 1` and even a plain +`unshare --map-root-user true` is refused) and passes on `nixos`. Lift +AppArmor's restriction for Nix builds there, persistently and as narrowly as +works, apply it when no gate is running, and check that the probe succeeds and +`just ci` passes `check flake`. diff --git a/quest/m1/bbr-collapse-estimate.md b/quest/m1/bbr-collapse-estimate.md index 676ea5aa7ea..b921a255bca 100644 --- a/quest/m1/bbr-collapse-estimate.md +++ b/quest/m1/bbr-collapse-estimate.md @@ -18,8 +18,9 @@ Evidence from #7671's sparse harness: after the path fell from 2 Mbit/s to 100 kbit/s, the estimate held 1.32 Mbit/s for about 11 s, and in the FFI run the top rung disabled 5.6 s after the 500 kbit/s step. The probe idles while the estimate covers the target, and the reported estimate -(`max_bw.min(bw_delay)`) ignores the loss bound BBR3 keeps -(`inflight_longterm` and its short-term bounds in -`rs/moq-congestion/src/bbr3/mod.rs`). Measure first, change second, and -record both runs. Keep the change inside the reported estimate unless the -measurement shows the control loop itself is late. +(`max_bw.min(bw_delay)` in `metrics`, `rs/moq-congestion/src/bbr3/mod.rs`) +ignores the loss bounds BBR3 keeps: `bw_shortterm`, which +`bound_bw_for_model` already folds into the model's `bw`, and the inflight +bounds `inflight_longterm` and `inflight_shortterm`. Measure first, change +second, and record both runs. Keep the change inside the reported estimate +unless the measurement shows the control loop itself is late. diff --git a/quest/m2/watch-symlink-targets.md b/quest/m2/watch-symlink-targets.md index 8e5a6150dc9..26c545913c6 100644 --- a/quest/m2/watch-symlink-targets.md +++ b/quest/m2/watch-symlink-targets.md @@ -2,9 +2,9 @@ ## Goal -Rewriting a certificate or key in place reloads it even when the configured -path is a symlink whose target lives in another directory, as Kubernetes -secret mounts do. This covers roots, served chains, and client certificates. +Rewriting a certificate or key in place reloads it when the configured path +is a symlink whose target lives outside the configured file's directory. This +covers roots, served chains, and client certificates. ## Plan @@ -12,7 +12,9 @@ Decided 2026-11-05 (maintainer, follow-up from [moq-dev/moq#7906](https://github.com/moq-dev/moq/pull/7906)): ✅ m2. `watch::callback` (`rs/moq-tokio/src/watch.rs`) watches each configured -file's directory, so a rewrite of a symlink target elsewhere goes unseen. -Also watch the directories the configured paths resolve to, and re-resolve -them when a link changes, so a swapped `..data` link is followed. Tests drive -the watcher with controlled events, not wall-clock waits. +file's directory. Its filter already reloads when a path resolves somewhere +new, which covers a Kubernetes `..data` swap and a repointed link, but an +in-place write to a target in an unwatched directory raises no event. Also +watch the directories the configured paths resolve to, and move those watches +when a resolution changes. Tests drive the watcher with controlled events, not +wall-clock waits.