fix(compile): rewrite backslashes on the PIO no-compile-CWD fallback path - #912

… (#3476)

Lands the real-hardware `I2sPeripheralEsp32DevEsp` class that
implements `II2sPeripheralEsp32Dev` against IDF v4's `driver/i2s.h`
DMA + ISR machinery. Completes the peripheral surface that was
scaffolded in #3473 (Stage 1).

## What Stage 2 delivers

- `i2s_peripheral_esp32dev_esp.h` — declares the real-hw class
  paralleling the mock. Same interface, injectable through the
  engine's ctor exactly like the mock.
- `i2s_peripheral_esp32dev_esp.cpp.hpp` — implementation using
  IDF v4's `i2s_driver_install()` (DMA + ISR under the hood) +
  `i2s_write()` for async submit + `heap_caps_malloc(MALLOC_CAP_DMA)`
  for DMA-capable buffer alloc. Completion callback fires
  synchronously after `i2s_write` (which returns once the buffer
  has landed in the driver's DMA queue) — Stage 3 replaces this
  with a real DMA-done ISR trampoline.

## Deliberately not linked yet

The new `.cpp.hpp` is intentionally NOT included in the classic-
ESP32 `_build.cpp.hpp` unity build. Its `#include "driver/i2s.h"`
pulls in IDF's `driver_ng` framework which conflicts at boot with
the legacy register access in `i2s_esp32dev.cpp.hpp` (Yves Bazin's
long-standing driver), producing:

    E (586) ADC: CONFLICT! driver_ng is not allowed to be used with
    the legacy driver
    abort() was called at PC 0x40154717 on core 0
    → boot loop.

Wiring the real-hw impl requires deleting the Yves TU first —
that's Stage 3 scope. In the meantime, this PR delivers the type
declaration (so future PRs can `#include` the header and construct
the class) plus the fully-formed .cpp.hpp waiting to be linked
once the Yves cleanup lands.

## Verification

- `bash compile esp32dev --examples AutoResearch` — succeeds
  (fbuild fix for the PIO source-path backslash regression landed
  separately in FastLED/fbuild#912; without that fix even master
  was blocked on this box).
- Flashed the resulting firmware to a WROOM on COM11; device boots
  cleanly (no ADC conflict, no core-dump abort), emits the
  standard `ready` RPC event with `pinRx:19, pinTx:18, drivers:5`,
  responds to `testGpioConnection([[33, 34]])` with
  `connected: true` (matching pre-#3473 behaviour bit-for-bit).
- `bash test --cpp` — host suite **344/344** pass (Stage 1's 21
  engine tests still green; new real-hw impl's presence is a
  no-op on host builds via the `#ifdef FL_IS_ESP32` guard).
- `clang-format --dry-run --Werror` clean on every touched file.

## Follow-up (Stage 3, tracked separately)

- Delete `i2s_esp32dev.{h,cpp.hpp}` + `clockless_i2s_esp32.{h,cpp.hpp}`
  (Yves Bazin's TU) so the driver_ng conflict clears.
- Include `i2s_peripheral_esp32dev_esp.cpp.hpp` in the `_build.cpp.hpp`.
- Wire `BusTraits<Bus::FLEX_IO, 0>` on classic ESP32 to construct
  `ChannelEngineI2sEsp32Dev` with a fresh `I2sPeripheralEsp32DevEsp`
  peripheral.
- Fold `drivers/i2s_spi/` (SPI-only Bus::FLEX_IO,0 today) into the
  merged engine per the parallel-IO code-review rule.
- Replace the sync completion callback with an ISR trampoline.
- Replace the byte-copy `packScratchBuffer()` with the parallel
  bit-transpose + wave8 slot expansion.
- WROOM on-device WS2812B decode via `bash autoresearch --i2s`.

Ref #3474.

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

The #618 sync module and #912's backslash-rewrite fix were combined
into a single squash-merge (ecd80c7). Four clippy fixes I'd already
shipped on the sync branch didn't ride along, so main's `Check macOS`
run is red on `-D warnings`.

Reapplying:

- fbuild-serial/src/boards.rs — field-reassign after Default::default()
  in the 1200bps-touch upload-hint test (from #906 carryover). Use
  struct-literal syntax + ..Default::default().
- fbuild-cli/src/sync/source.rs — swap the closure form
  `.rsplit(|c| c == '/' || c == '\')` for the array-pattern form
  `.rsplit(['/', '\'])`. Same behavior, satisfies
  clippy::manual_pattern_char_comparison.
- fbuild-cli/src/sync/mod.rs — fold the nested `if !prompt_multi_env(..)`
  into the outer condition (clippy::collapsible_if). Same behavior.
- fbuild-cli/src/cli/sync_cmd.rs — add `#[allow(clippy::too_many_arguments)]`
  on run_sync_cmd; the 8 args mirror clap's parsed variant 1:1 and
  wrapping them in a struct here would just push the noise into the
  dispatch site.

No behavior changes.

…#1372)

* ci(dylint): add a Windows leg so Windows-only code is actually linted

Closes #1359.

`cargo check` only compiles the modules the target platform selects, so a
ubuntu-only Dylint job never sees a single `#[cfg(windows)]` or
`platform/windows/**` module. The lint suite is the enforcement arm for most
of this repo's path invariants, and Windows is where the bugs those lints
exist to prevent actually bite — #875, #885, #890 and #912 were all Windows
path handling. The gate was pointed away from the platform that needs it.

Not hypothetical. A full sweep on a Windows host against main reports:

    error: use fbuild_core::path::NormalizedPath::display_slash() instead of
           hand-rolled backslash-to-slash rewrite
      --> crates\fbuild-core\src\platform\windows\fs.rs:21:17

## Fixing the violation, not excusing it

`ban_manual_slash_normalize` allowlisted `fbuild-core/src/path.rs` as "the
primitive itself". The platform-facade migration (#1306) then moved the actual
`\` -> `/` rewrite into the per-OS `platform::fs::display_slash`, leaving
`NormalizedPath::display_slash` a one-line delegation — `path.rs` no longer
performs the transformation at all.

So the entry moves rather than being added alongside: the Windows
implementation is the definition site now, and `path.rs` is dropped from this
lint's allowlist. Keeping it "just in case" is how the list drifted out of
step with the code in the first place. `path.rs` keeps its entries on the
*other* lints, which it still earns.

Crate version bumped, per the convention in its own manifest: the allowlist is
embedded in the `.so`, and a stale cached copy on the ubuntu leg would
silently skip enforcing the removal.

## The workflow change

A matrix with per-leg pinned `name`, not a derived one. `Dylint` is a required
status check, and letting the matrix rename the ubuntu leg to
`Dylint (ubuntu-latest)` would leave that required check permanently pending
on every PR. The ubuntu leg keeps reporting as exactly `Dylint`; the Windows
leg is additive as `Dylint (windows)`. `fail-fast: false`, so a Windows
violation cannot mask a Linux one.

`defaults.run.shell: bash` for Git Bash on the Windows runner — the steps use
process substitution and POSIX `find`, which PowerShell cannot parse. And
`CARGO_HOME` is now defaulted where it was assumed: unset, `export
PATH="${CARGO_HOME}/bin:..."` would silently prepend a bare `/bin`.

## Verified on Windows, which is the new leg

Every step of the job was run locally on a Windows host:

- full `dylint --all -- --workspace --all-targets` sweep: **0 violations**
- the observed-comparison that closes the job
  (`enforce_platform_boundary.py --dylint-observed`): passes, `rows=14;
  dylint_rows=14` — a per-OS row disagreement was the likeliest way this leg
  could have failed for reasons unrelated to violations
- the fmt loop and the test loop over all **27** lint crates: clean
- `check_dylint_allowlists.py`, `render_workflows.py --check`, and
  `check_workflow_concurrency.py`: pass

## Not included: a macOS leg

Same one-line matrix addition, deliberately deferred. `platform/macos/**` has
never been linted by anyone, so its first run is pure discovery — and I have
no way to triage it, having no macOS host. Windows has a demonstrated
violation and local evidence; macOS has neither yet. Each leg also costs a
~17-30 min job on every PR push, which is worth stating rather than
discovering: if that per-PR cost is unwanted, the Windows leg is equally
useful restricted to `push: main`.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* ci(dylint): scope the lint crates' own tests to the ubuntu leg

The new Windows leg went red on its first run, and the finding is worth
recording because it is not what the leg exists to catch.

The lints themselves are clean on Windows — the workspace sweep passes. What
failed is the "Test Dylint libraries" step, which runs each lint crate's own
compiletest ui fixtures:

    error: could not load library
    `.../target/dylint-tests/debug/ban_manual_slash_normalize@nightly-2026-04-16.dll`:
    LoadLibraryExW failed

Dylint 6.0.1 looks for its test library under `<target>/debug` while soldr
sets `CARGO_BUILD_TARGET`, so cargo writes to `<target>/<host>/debug`. Each
lint's `fn ui` already clears that variable — that is what the comment in
every one of them is about — but on the runner the load still fails. A
dylint/soldr/compiletest interaction, not an fbuild defect.

Rather than work around it, the step is scoped to ubuntu, and the scoping is
principled rather than a dodge: those are the lint crates' *own* tests,
asserting each lint fires on its fixture. That behavior is platform-
independent and already covered once. What the Windows leg uniquely provides
is compiling Windows-gated *workspace* source so the lints can see it — which
is the entire point of #1359, and which passes. The same reasoning scopes the
fmt loop and the three Python validators, none of which learn anything from a
second OS.

Side benefit: the Windows leg gets ~10 minutes shorter, which matters for a
job that runs on every PR push.

The Windows leg now runs exactly the steps verified locally on a Windows host:
the full sweep (0 violations) and the observed-comparison (`rows=14;
dylint_rows=14`).

Refs #1359

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* ci(dylint): pin the workflow token to contents: read

CodeRabbit: the workflow had no `permissions` block, so it inherited the
repository/organization default, which can include write access.

Every step in this job reads — checkout, a shallow `git fetch origin main` for
the shrink-only allowlist diff, and the lint runs themselves. Nothing writes,
so nothing should be able to.

Worth noting this is not yet the repo-wide convention: 8 of 106 workflows set
`permissions` today. Hardening the rest is a worthwhile sweep but belongs in
its own change rather than riding along here.

Refs #1359

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>