Closes the orchestrator gap exposed by aethersdr/AetherSDR#2770:
when the agent moves a symbol from one translation unit to another
(or replaces a local declaration with a reference to one in a shared
module), the main app keeps linking cleanly because it pulls in
the full CORE/MODEL/GUI source lists. But unit-test targets that
link a minimal subset fail with `undefined reference to ...` —
caught only by CI after a full build cycle, requiring a maintainer
patch on top.

New step 6 (between Implement and Commit) mandates a grep + local
build of every affected test target before commit. Calls out the
transitive-dependency case (LogManager.cpp → AsyncLogWriter.cpp →
AppSettings.cpp) that caught #2770 specifically, and gives the
agent a heuristic shortcut (the words "consolidate", "extract",
"share", "centralize", "deduplicate" → always run the audit).

The existing step 6 (commit-message construction with constitution
principle citation) becomes step 7. No other behavioral change.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Covers the matrix from the issue: disabled hot-path no-op, window
aggregation count/p95, window-reset isolation, the full stall-threshold
matrix (panUpdate / waterfallUpdate / render / frameAge / udpDrain /
input), drag-aware UI lag thresholds (50 ms idle / 33 ms drag), the
frame-restart counter, percentile-95 boundary cases (empty / single /
[1..100] sorted / 100 reversed), and the 1 s window cadence (no fire
at 950 ms, fire at 1100 ms).

Two purely additive test seams land on PerfTelemetry:
`setClockOverrideForTest(qint64)` pins nowNs() to a synthetic value so
window timing is deterministic without real sleeps, and `resetForTest()`
zeroes aggregation state between cases so the singleton doesn't leak
between tests. The clock override is a single relaxed atomic load in
the hot path; when zero (production default) the original steady_clock
path is used unchanged, so the existing record* call surface that the
17 high-risk callers reach is untouched in shape or behavior.

The test executable links LogManager.cpp + AsyncLogWriter.cpp +
AppSettings.cpp instead of redeclaring its own Q_LOGGING_CATEGORY for
`aether.perf` — this honors Principle III's consolidation (one
canonical definition site per category, per #2770) and matches the
linkage pattern established by `ax25_libmodem_shim_test`. Output is
captured via qInstallMessageHandler filtered to `aether.perf` and
matched on key=value field extraction from emitted PerfStall and
PerfSummary lines.

Blast radius: risk_score=0.418, 17 high-risk affected (top:
MainWindow::MainWindow, MainWindow::wirePanadapter, MainWindow::buildUI,
RadioModel::onStatusReceived, MainWindow::buildMenuBar). All these
reach PerfTelemetry through its existing public record* API which is
unchanged; the additions are two new public test-only methods plus one
relaxed atomic load in nowNs() that is a no-op when the override is
the default zero.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

## Summary

Fixes #2500

### What was changed

Covers the matrix from the issue: disabled hot-path no-op, window
aggregation count/p95, window-reset isolation, the full stall-threshold
matrix (panUpdate / waterfallUpdate / render / frameAge / udpDrain /
input), drag-aware UI lag thresholds (50 ms idle / 33 ms drag), the
frame-restart counter, percentile-95 boundary cases (empty / single /
[1..100] sorted / 100 reversed), and the 1 s window cadence (no fire
at 950 ms, fire at 1100 ms).

Two purely additive test seams land on PerfTelemetry:
`setClockOverrideForTest(qint64)` pins nowNs() to a synthetic value so
window timing is deterministic without real sleeps, and `resetForTest()`
zeroes aggregation state between cases so the singleton doesn't leak
between tests. The clock override is a single relaxed atomic load in
the hot path; when zero (production default) the original steady_clock
path is used unchanged, so the existing record* call surface that the
17 high-risk callers reach is untouched in shape or behavior.

The test executable links LogManager.cpp + AsyncLogWriter.cpp +
AppSettings.cpp instead of redeclaring its own Q_LOGGING_CATEGORY for
`aether.perf` — this honors Principle III's consolidation (one
canonical definition site per category, per #2770) and matches the
linkage pattern established by `ax25_libmodem_shim_test`. Output is
captured via qInstallMessageHandler filtered to `aether.perf` and
matched on key=value field extraction from emitted PerfStall and
PerfSummary lines.

Blast radius: risk_score=0.418, 17 high-risk affected (top:
MainWindow::MainWindow, MainWindow::wirePanadapter, MainWindow::buildUI,
RadioModel::onStatusReceived, MainWindow::buildMenuBar). All these
reach PerfTelemetry through its existing public record* API which is
unchanged; the additions are two new public test-only methods plus one
relaxed atomic load in nowNs() that is a no-op when the override is
the default zero.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

### Recovery note

This PR is being opened manually. The aetherclaude orchestrator (trace
fe2b51d0) pushed the signed commit at 17:53 PDT but died at the next
bash line due to a set -e race with commit-signed.js's exit code. Root
cause + fix shipped as 4832725 (run-agent: guard commit-signed +
create-pr against set -e silent death).

---
Generated by AetherClaude (automated agent for AetherSDR) — PR creation
completed manually after orchestrator interrupt.

---------

Co-authored-by: aethersdr-agent[bot] <273844287+aethersdr-agent[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>

…ust/#2780) (#2786)

## Summary

Cherry-pick of @rfoust's panadapter persistence and macOS GPU lifecycle
work from PR #2780, applied cleanly onto current main with the
stale-snapshot reverts left out.

## Why a cherry-pick rather than merge of #2780

PR #2780 was based on commit \`a0512d99\` (PR #2772 merge), and its
single Codex-generated commit included **stale-snapshot content for
files outside the intended scope**. Squash-merging #2780 as-is would
have silently reverted two recently-landed PRs:

- **PR #2770** (\`aether.ax25\` Q_LOGGING_CATEGORY consolidation) —
\`AetherAx25LibmodemShim.cpp\`, \`Ax25HfPacketDecodeDialog.cpp\`, plus
the test-target linkage in \`CMakeLists.txt\`
- **PR #2772** (RadioSetupDialog → PersistentDialog migration) —
\`RadioSetupDialog.{cpp,h}\`

Verified by three-way merge dry run: post-merge state showed \`class
RadioSetupDialog : public QDialog\` (pre-#2772) and 5 occurrences of
\`lcAetherAx25Shim\` (pre-#2770).

The actual panadapter work in PR #2780 is real and valuable, so this PR
extracts it onto a fresh branch from current main and skips the
unintentional reverts.

## What's in this PR (= the intended scope of #2780)

Cherry-picked verbatim from #2780:

| File | Change |
|---|---|
| \`src/gui/MainWindow.{cpp,h}\` | Persist/restore PanadapterLayout +
FloatingPanIds; expand layout count handling from 4 to 8 pans; preserve
unseen floating IDs across reconnect |
| \`src/gui/PanadapterStack.{cpp,h}\` | Sensible-fallback
docked-splitter layouts when some pans are floating or saved layout
count mismatches |
| \`src/gui/SpectrumWidget.{cpp,h}\` | macOS QRhiWidget lifecycle:
\`WindowAboutToChangeInternal\` before reparent; QRhi resource release
during shutdown while parent backing stores are valid; guarded late
callbacks; auto FFT floor relock after geometry/ypixels changes |
| \`src/core/PanadapterStream.cpp\` | Defensive dBm-range clamping (-180
floor, ≥10 dB span); FFT decode clamps pixel input and dbm output to
range |
| \`src/models/RadioModel.cpp\` | Use actual FFT pane height for radio
\`ypixels\` rather than total widget height |

Net: **+548 / -215** across 8 files.

## What was NOT cherry-picked (stale-snapshot reverts)

| File | Reason |
|---|---|
| \`CMakeLists.txt\` | -7 lines reverting #2770's test-target linkage I
added to fix the link error |
| \`src/core/tnc/AetherAx25LibmodemShim.cpp\` | Robbie's content has
local \`Q_LOGGING_CATEGORY(lcAetherAx25Shim, ...)\` — pre-#2770. Main
has the consolidated \`lcAx25\` from LogManager |
| \`src/gui/Ax25HfPacketDecodeDialog.cpp\` | Same pattern —
\`lcAetherAx25Dialog\` local, pre-#2770 |
| \`src/gui/RadioSetupDialog.{cpp,h}\` | Robbie's diff was **100% pure
revert** of #2772 — no new content, no panadapter-related additions.
Confirmed by inspecting the full \`git diff a0512d9..e72d7e4 --
RadioSetupDialog.cpp\` output |

## Attribution

Commit author is set to **Robbie Foust** so his contribution shows in
\`git log\` and the contributor stats. PR #2780 is being reset to
**Draft** (not closed) so Robbie has a clean reference point + can
update with fresh work if he wants.

## Test plan

- [x] Local build clean (402/402 link, only pre-existing warnings)
- [x] Three-way merge dry-run against current main: clean
- [ ] Manual on Linux: multi-pan layout, detached panadapter, restart,
confirm layout + detached state restore
- [ ] Manual on macOS GPU build: two detached panadapter windows, Cmd-Q,
restart, confirm both restore without the QRhi cleanup callback crash
- [ ] Manual: detach/redock panadapter and confirm auto FFT floor
relocks

## Cross-references

- Replaces #2780 (will be set to Draft after this PR opens)
- Same stale-snapshot pattern as aetherclaude #34 (Claude Code), but
from OpenAI Codex this time — worth a heads-up to whoever's watching
agent-output quality

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Robbie Foust <rfoust@duke.edu>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>