Effects: Flanger on audiodelays.Flanger, and a round of fixes across the library

.github/workflows/tests.yml

@@ -14,6 +14,9 @@ on:
     branches: [main]
     paths-ignore: ['docs/**', 'README.md', 'ROADMAP.md', 'CHANGELOG.md', 'AGENTS.md', 'LICENSE']
   pull_request:
+    # `labeled` too: adding the `lifecycle: all classes` label re-runs the
+    # pull request with the whole lifecycle matrix (see the test step).
+    types: [opened, synchronize, reopened, labeled]
     paths-ignore: ['docs/**', 'README.md', 'ROADMAP.md', 'CHANGELOG.md', 'AGENTS.md', 'LICENSE']
   # The full matrix on demand, for a branch that should see every Python
   # before it merges rather than after.
@@ -47,8 +50,30 @@ jobs:
           - os: macos-latest
             python: "3.13"
     runs-on: ${{ matrix.os }}
+    env:
+      # The lifecycle matrices (quick mode). All 45 classes take about 4.4
+      # CPU hours, over an hour on this runner, and the 13 audiodsp nodes
+      # under them half an hour more on one worker, so a pull request runs
+      # only what its change can move: the class or node whose own code
+      # changed, and when a shared file changed (the base class, the
+      # matrix, the pin) a fixed sample of each, nine classes in about 13
+      # minutes and four nodes in about five (SAMPLE in
+      # tests/test_lifecycle_matrix.py, NODE_SAMPLE in
+      # tests/test_node_matrix.py). A push to main runs every class and
+      # every node, and so does a manual run (workflow_dispatch) or a pull
+      # request labelled `lifecycle: all classes`.
+      LIFECYCLE_CHANGED_FROM: "${{ github.event_name == 'pull_request' && !contains(github.event.pull_request.labels.*.name, 'lifecycle: all classes') && format('origin/{0}', github.base_ref) || '' }}"
     steps:
+      # A pull request fetches its history, so the matrices can ask git what
+      # the change touches against origin/<base> (LIFECYCLE_CHANGED_FROM).
+      # Written `!= ... && 1 || 0` because `== ... && 0 || 1` is always 1:
+      # 0 is false in an expression, so the `||` took over and every pull
+      # request checked out at depth 1. With no base to diff from, the
+      # matrix then fell back to all 45 classes, and a pull request ran
+      # 1 h 43 min. A base git cannot diff from is now an error instead.
       - uses: actions/checkout@v7
+        with:
+          fetch-depth: ${{ github.event_name != 'pull_request' && 1 || 0 }}
       - uses: actions/setup-python@v7
         with:
           python-version: ${{ matrix.python }}
@@ -109,6 +134,17 @@ jobs:
       # a test. The tests would generate them too; this makes it explicit.
       - run: python tools/effect_probes/make_probes.py --ensure
 
+      # What the two matrices will run, in the log where a reader can check
+      # it, and a scope git cannot work out fails here by name.
+      - name: Show the lifecycle scope
+        shell: bash
+        run: >-
+          python -c "from tests.test_lifecycle_matrix import classes;
+          from tests.test_node_matrix import names;
+          print('base:', '${LIFECYCLE_CHANGED_FROM:-none, all}');
+          print('classes:', ' '.join(classes()) or 'none');
+          print('nodes:', ' '.join(names()) or 'none')"
+
       # The same tests `python -m unittest discover -s tests -p "test_*.py"`
       # finds, spread across every core the runner has: serially the suite
       # takes over an hour. A test that writes a file writes it into its own

AUDIODSP_PIN

@@ -345,5 +345,57 @@
 # holds its line still at mix 0 (#226); audiovorbis (#238, #239); and an
 # fmod-free angle reduction for the RP2040 (#225).
 #
+# 2026-10-13: v0.6.6rc1 -> f7a86c9, the effects program's Phase 7 node batch
+# (step 7b): the candidate for audiodsp's next pre-release, v0.6.6rc3, and
+# the final pin of the phase. Named by commit until the tag is cut; the tag
+# names the same source. What moves beneath the components, by `git log
+# v0.6.6rc1..f7a86c9 -- src`: audiodelays.Flanger lets go of a looped source
+# that comes back empty and gives a stalled one silence, instead of spinning
+# the pull; audioecho.FeedbackDelay's upward loop_semitones writes its line
+# through a band limit at 0.7 of the frequency the next pass would carry to
+# Nyquist, so a climb fades out at the top instead of folding back; and both
+# shifters (the loop's two taps and audiodelays.PitchShift's splice) move a
+# splice's incoming side into phase with the outgoing one, so a held note
+# plays its shifted pitch, and a chord its notes, instead of cancelling.
+# Unshifted and downward FeedbackDelay renders, silence and a lone click
+# through either shifter, and a frozen PitchShift window render the bytes they
+# did.
+#
+# Moved to v0.6.6rc3 (a9d36e6) on 2026-10-13: the release of the commit
+# above (audiodsp#245, #247), the same node tree as f7a86c9 outside VERSION
+# and CHANGELOG.md, so CI can install it from GitHub.
+#
+#
+# Moved to v0.6.6rc4 (8d5c282) on 2026-10-14: MultiTapDelay's reset puts the
+# write heads back where a fresh node has them, and audioconvolve.Convolver
+# gains wet_gain (default 1.0, byte-identical there), which
+# ConvolutionReverb uses to trim its four darkest rooms (audiodsp#248).
+#
+# Moved to v0.6.6rc5 (b278d32) on 2026-10-14: audioconvolve.Convolver gains
+# room_peak_gain and room_peak_gain_for, the reading ConvolutionReverb's wet
+# trim is computed from. Every render is byte-identical to rc4's.
+#
+# Moved to v0.6.6rc6 (9399602) on 2026-10-14: audioecho.FeedbackDelay gains
+# wow_depth_ramp_ms, which Vibrato's Engage rides so its Rise ramps on every
+# interpreter, and audioshaper.Waveshaper gains recentre(), which
+# Saturation calls so a gain move in silence with Hysteresis on plays
+# nothing. Every render is byte-identical to rc5's at the defaults of both.
+#
+# Moved to v0.6.6rc8 (f83ddc6) on 2026-10-14, by way of rc7: the CPython
+# extension no longer crashes in garbage collection while a RawSample still
+# exports a buffer; audiodynamics.Dynamics writes a mono block at its own
+# stride, so short source buffers leave no stale frames; audioroute.Splitter
+# gains clear(), which six classes call to drop what the ring held from
+# before a bypass when Mix comes back up; and audioecho.FeedbackDelay takes
+# a new wow_shape by dipping the depth to zero and swapping there, which
+# Vibrato hands its clock law to. Every other render is byte-identical to
+# rc6's.
+#
+# Moved to v0.6.6rc9 (96abd34) on 2026-10-15: audiobiquad.Biquad walks a
+# coefficient move over 20 ms instead of keeping its state through the jump,
+# so LowPass, BandPass and Notch no longer thump at full scale when
+# Frequency jumps to the bottom (audiocomponents#162). Static renders are
+# byte-identical to rc8's; only renders across a coefficient move change.
+#
 # <ref> <commit>   -- ref is what CI installs (git+https://...@<ref>)
-v0.6.6rc1 30cbbf619751ab6ff7c40d99336907af0a081fa6
+v0.6.6rc9 96abd345344eb356f1af5de24e320bfcf778cb8d