← board

No wave module — and TSP's import of it is unguarded

Row 2 of devdocs/dev/tsp-compile-wall-inventory-2026-09-20.md. Measured 2026-09-20 against TSP's live tree a143976 and pxx 45b8571d7: of the 26 distinct modules TSP imports, 23 resolve; the three that do not are ctypes, __pxx__ and wave.

Why this one is worth taking on its own

The other two are one story — the platform seam, which is blocked on application code that does not exist (tsp/platform/_pxx_backend.py), and on the owner. wave is none of that. It is ours, it is small, it is plain stdlib, and tsp/voice.py:29 imports it unguarded:

import wave

So there is no try:/except ImportError: to fold away at compile time and no application edit that avoids it. It blocks whether or not the seam is ever ported.

The surface TSP actually uses — six calls, both directions

tsp/voice.py:205-224, scaled_copy(), which rewrites a 16-bit WAV at a lower volume:

with wave.open(path, "rb") as w:
    params = w.getparams()
    data = array.array("h", w.readframes(w.getnframes()))
...
with wave.open(tmp, "wb") as w:
    w.setparams(params)
    w.writeframes(data.tobytes())

That is: wave.open(path, mode) returning an object usable as a context manager, getparams() / setparams() round-tripping the header tuple (nchannels, sampwidth, framerate, nframes, comptype, compname), getnframes(), readframes(n) → bytes, writeframes(bytes). Uncompressed PCM only; TSP never touches the compressed arms and CPython's wave barely supports them either.

Underneath it is RIFF chunk walking — a 12-byte header, a fmt chunk and a data chunk. No codec.

array is already present as lib/rtl/mimic_array.pas, so the array.array("h", ...) on both sides of this function is not an additional gap.

Shape of the fix

Per devdocs/dev/pxx-crash-course.md: a lib/rtl unit carries a Pascal AND a Python surface in ONE unit — this does not need a separate mimic_ wrapper if a Pascal WAV reader/writer is written with the Python-shaped entries beside it. Check first whether one already exists under another name (lib/rtl/image.pas is the precedent for "media format, both surfaces, one unit"); grep before writing, this tree has had a library reported as missing while it was present.

What would retire this

tsp/voice.py compiling, and a fixture that writes a WAV and reads it back with the params round-tripping — including a sampwidth other than 2, so the expected value is not also the default.

What this ticket does NOT claim

The census instrument reports the FIRST error per file, so clearing wave is not promised to deliver voice.py: the module also uses subprocess, threading and a generator expression into array.array, any of which may wall behind it. The count is a direction, not a size.

RESOLVED 2026-09-20 — lib/rtl/mimic_wave.pas

Uncompressed RIFF/WAVE, read and write, resolving through the mimic_ fallback (note: wave -> mimic_wave (shim, subset)). No compiler change: the resolver already did its half, there was simply no unit behind the name.

Surface implemented: open/openfp, the context-manager protocol, the six getters plus getparams, the six setters plus setparams, readframes, writeframes/writeframesraw, rewind, tell, setpos, close, and wave.Error. Wave_read/Wave_write are aliases of one class — a Pascal function has one return type, so two classes would need open to return a common base, and that base is this class. Absent on purpose, and said in the unit: markers, reading from an open file object, iteration.

Verification

test/lib_mimic_wave.npy, 63 rows, a DIFFERENTIAL that runs unmodified under CPython — every expectation read off the oracle, not typed from the RIFF documentation. Green in three places: CPython 3.14.4, pxx at HEAD, and under the pinned compiler, so the row is not inert waiting for a make pin. The file pxx writes is byte-identical to the one CPython's own wave writes for the same input, and CPython reads pxx's file back with the right values.

THE NEGATIVE CONTROL IS THE PART WORTH READING

At 61 rows the test was green under CPython, green at HEAD, green under the pin, and byte-identical to the oracle's output. I then broke the shim deliberately in two places and 61 of 61 stayed green.

Two rows added — pad.byterate on the stereo clip, and the hand-built LIST chunk resized to seven bytes. Both are red on the broken build and green on the real one. Generalised to the playbook: two quantities that are equal in your fixture cannot both be tested by it.

WHAT THIS DID NOT DELIVER, which the ticket predicted

tsp/voice.py does NOT compile. With --threadsafe, its wall moves from wave to threading.Condition at line 79 — no member Condition came of the qualifier threading. That is the RTLEvent/threading hole (feature-b-the-rtlevent-family-is-absent-from-the-threading-rtl), a different lane and a different ticket. The wall is cleared; the unit is not delivered, exactly as this ticket said would happen.

One correction to this ticket's own framing, for the record: wave was found by the MODULE-SURFACE scan (26 imported, 23 resolve), not by the first-error census — voice.py's first error was always threading at line 26, three lines above the wave import. The census never named wave and could not have. That is the first-failure blindness this tree keeps recording, working in the useful direction for once: the surface scan sees walls the census cannot.

Log