spu: reverb, noise, sweeps and SPU IRQ (closes #103); CAUSE.IP2 combinational (supersedes #102)

[mstan]

Closes #103. Absorbs and supersedes #102.

Implements the four SPU gaps Issue #103 names — reverb, noise generator, volume
sweeps and SPU IRQ — and lands the separable, independently verified parts of
PR #102. Full per-piece disposition, including everything I could not settle,
is in docs/internal/SPU_FIDELITY_103.md.

Why this supersedes #102 rather than merging alongside it

#102 covers 1 of #103's 4 items. I verified that by grepping its complete
diff for reverb|noise|sweep|LFSR|vLOUT|vROUT|dAPF — zero matches. It
implements SPU IRQ plus three unrelated fixes. Its SPU IRQ and this branch's
would collide textually in spu.c, so there is one implementation, here.

Everything of Alexbeav's that stands is carried, and he is credited as
Co-Authored-By on the implementation commit:

• CAUSE.IP2 combinational — his central claim, and it is correct.
  Checked against the in-tree Beetle oracle rather than accepted on reasoning:
  beetle-psx/mednafen/psx/irq.cpp recomputes the line as
  AssertIRQ(0, Status & Mask) at power-on, at every assert, and at both the
  status-ack and mask-write halves of a register write, and cpu.cpp's
  AssertIRQ clears CAUSE bit 10 unconditionally before re-setting it from
  the level. That is a combinational mirror, and it matches R3000A, where the
  Cause.IP field is not storage.
• SPU RAM DMA readback — DMA4 in the SPU→RAM direction wrote literal zeros.
  SPU RAM is readable memory; that was not a transfer at all.
• Mid-dispatch audio pump — the SPU is autonomous on hardware, so a guest
  busy-wait that never presents a frame must not freeze SPU time.

Two defects in #102 are fixed here:

1. A second latching IP2 writer it missed. runtime/src/psx_interpreter.c
   had its own cop0[COP0_CAUSE] |= (1u << 10) at delivery — the identical
   defect in the interpreter path. An audit of every cop0[13] writer confirms
   no other site touches bit 10, and psx_irq_refresh_cause_ip2() is now the
   sole owner.
2. The audio pump would have broken every turbo load. spu: implement SPU IRQ + CAUSE.IP2 combinational + mid-dispatch audio pump #102 wires the hook
   to sdl_audio_pump() unconditionally, but sdl_audio_update() is the sole
   authority on whether to emit, discard, or not pump at all — during a
   turbo-load hard mute it deliberately does not pump, so the queue drains and
   voice positions freeze in place and music resumes where it left off.
   Pumping unconditionally from the VBlank edge bypasses that and would replay
   time-compressed audio on every load. The hook now goes through a wrapper
   that mirrors the gate state.

#102's own validation (Medal of Honor Underground, Gran Turismo) could not
be reproduced — neither disc is in our suite — so those claims are not repeated
here as if verified. What is verified is the hardware behaviour behind each
change, plus non-regression on our own titles.

Clean-room, and why that mattered

psxrecomp ships under PolyForm Noncommercial 1.0.0; Beetle PSX/mednafen is
GPL-2.0-or-later. Incompatible. Beetle's tree contains a complete, correct
reverb, so the obvious route is closed, and two contributors walked into it:
PR #16 was parked on license grounds after an audit found it tracked
beetle-psx/mednafen/psx/spu.cpp lines 590–730, and PR #13's reverb was
separately wrong — it gated on bit 15 of dAPF1, an address offset whose top
bit is set for any offset ≥ 0x8000, so reverb switched on and off according to
an address value. The real gate is SPUCNT bit 7.

This is written from the psx-spx / nocash register map and documented
algorithm. Beetle's source was not consulted for the implementation; it is used
only as a runtime output oracle, which is not a derivative work. That is the
path docs/internal/upstream/martin-pr16-spu-reverb.md prescribes.

What landed

• Reverb — documented 22050 Hz engine: vLIN/vRIN input, same-side and
  crossed different-side IIR reflections, 4-tap comb early echo, both all-pass
  filters, vLOUT/vROUT output, work-area advance. SPUCNT bit 7 gates
  writes only; reads, IRQ checks, address advance and output always run, as
  on hardware. Addressing is provably confined to [mBASE, 0x80000). Per-voice
  sends from EON, CD send from SPUCNT bits 0+2.
• Noise — 16-bit LFSR with the documented parity feedback, clocked at the
  SPUCNT bits 8–13 rate; NON voices substitute it for their ADPCM sample.
• Volume sweeps — all 24 voice L/R plus main L/R, as live envelopes on the
  same rate machinery as ADSR; a guest read of a sweeping register returns the
  live level.
• SPU IRQ — checked at every SPU RAM access class (FIFO, DMA both
  directions, ADPCM block fetch, capture writes, reverb work-area accesses),
  gated on SPUCNT bit 6, mirrored in SPUSTAT bit 6, acked by clearing bit 6.
• Capture buffers — CD L/R and voice 1/3 written to SPU RAM, so a game that
  parks the IRQ address there still gets its interrupt.

The CD-only fast path in spu_render was removed rather than extended: the SPU
now has per-frame work that must run in every path, and an FMV (XA audio, zero
active voices) is exactly the CD-with-reverb case.

Verification

• test_spu_fidelity — 137 checks, 0 failures. New, from scratch. Covers
  IRQ on each access class plus ack/re-arm, reverb write-gating and work-area
  wrap confinement, the CD-only path, noise non-repetition, sweep rise/fall and
  live readback, direct-volume decode, and savestate round-trip.
• test_cause_ip2_combinational — rise, fall on ack, fall on mask, partial ack
  with another source pending, partial-width writes, preservation of
  ExcCode/BD/IP0/IP1. Both registered in CTest.
• BIOS LLE boot proves the reverb engine runs. The BIOS itself sets SPUCNT
  bit 7 and allocates a 62 KB work area at mBASE = 0x70940. Through the boot
  chime the work area holds 256/256 non-zero reflection halfwords peaking at
  24681, reverb_cur laps correctly inside the work area, and after the dry
  signal goes silent the work area keeps ringing down (628 → 18 → 2) for ~3 s —
  a decaying tail outliving its input. Noise LFSR observed shifting with correct
  parity feedback.
• Tomba 2 intro FMV, A/B against the exact parent commit with a fully warm
  shard cache and the pre-SPU exe as control: interp_share 0.7813 vs 0.7775,
  exc_share 0.2386 vs 0.2381, 39.1 vs 38.0–38.4 fps. Identical within noise —
  the added per-frame DSP work costs nothing measurable.
• MMX6, Rainy Turtloid stage, ~248 s at 59.9 fps with 36 host underruns
  total. Note honestly that this stage streams its audio as CD-XA and drives
  no SPU voices (active_mask = 0, eon = 0, non = 0, SPUCNT bit 7
  clear), so it validates no regression on the CD-only path — which is exactly
  where the fast path was removed, the CD bus made to drain continuously, and
  capture writes added — and not the reverb.

Known deviation, stated plainly

The reverb engine runs at 22050 Hz, but the hardware's 22.05 → 44.1 kHz
reconstruction filter is not specified in the documentation available under
the licensing constraint, and Beetle's filter was deliberately not copied. This
branch uses linear reconstruction, isolated in a single swappable function and
marked in the source. All ten places the documentation ran out are enumerated
in §6 of the design doc as an oracle-comparison work list. An
output-vs-Beetle audio comparison at a fixed scene is the natural follow-up and
has not been done.

Merge prerequisites

• recomp-ui must be bumped in lockstep, unrelated to this change but
  blocking: framework master's feat(mods): expose linked package attribution
  needs recomp-ui 99558ee, which adds author_links / author_link_count /
  source_name / source_url to RecompLauncherCModPackage. A title pinned to
  854ae10 fails to compile with 'RecompLauncherCModPackage' has no member named 'author_link_count'.
• Every title needs regen + revalidation, because CAUSE.IP2 changes core IRQ
  delivery for all of them. This is the highest-risk item in the branch. It is
  the faithful behaviour, so per Rule -1 it lands and titles are revalidated
  rather than the fix being narrowed.

Also recommend closing

feat/pr13-salvage-remainder retains nothing but the rejected reverb commit
5fc8e15d. Its other three commits are already on master — verified four ways
(git cherry -v marks them -; cherry-picking all three onto master produces
empty commits; and execute_ch5_pio, the CFC0 + Cause write-protection, and
the autocompile.c change are all present). So PR #13 can be closed and that
branch abandoned; origin/park/pr13-spu-reverb-rejected still retains the
commit.