[r-bin]

this is kind of part 2 of "why on earth does Evermore produce RNG with mid-frame memory" (I'd expect some minor merge conflicts with PR #278)
It used to be handled in src/soe_frame_driver.c, but that solution was very volatile. The recommended way for this to be merged was 3 small PRs that depended on each other, which I overruled for simplicity.

Summary

Five hardware timing rules that let a frame driver match MesenCE clock for clock. Each sits where its hardware lives, has its
own switch, and is on by default only with the master-clock APU timeline (RtlApuUseMasterClock, #241). Without that timeline
nothing changes.

 snes.c
-  every line 1364 clocks, every field 357,368
+  line 240 of every other non-interlaced field: 1360 clocks, no long dots   SNESRECOMP_SHORT_LINE
+    (Snes.fieldParity, toggled at V=0; savestate v12)
   $4212 read
-  bits 7, 0: from the beam's last sync
+  bits 7, 0: at the read's own clock, as bit 6 already is                   SNESRECOMP_HVBJOY_EXACT
 snes_refresh_charge
-  40 clocks per 1364 clocks of CPU time, anywhere
+  40 clocks per refresh point crossed, at H = 538 - (line start & 7)        SNESRECOMP_REFRESH_AT_BEAM
+  an APU port access behind a point sees the stall first
 snes_beam_frame_driver_run_frame
+  if SNESRECOMP_BEAM_EXACT: run_frame_exact (the old path is untouched)
+    CPU starts at master clock 186                                          SNESRECOMP_RESET_CLOCK
+    interrupt sequence: 8 cycles, 60 clocks, a refresh point stalls it
+    field NMI after the first instruction with a cycle from V=225 H=6       (snes/last_cycle.c)
+    NMI enabled inside vblank (RDNMI set): after the next instruction, same field
+    slices end at each refresh point; the beam follows the CPU after each
+    SnesBeamFrameHooks: frame_start, interrupt_entry, nmi_taken, field_nmi, slice_end

A host needs no timing code of its own for these: no beam alignment after a field, no walking the beam ahead for $4212, no
refresh top-ups. docs/FRAME_MODEL_HOSTS.md lists the switches.

Why

Each rule broke Secret of Evermore in a way a player or a replay sees; each touches code most SNES games contain.

Rule	What went wrong in Evermore without it	The pattern it affects in any game
Short line	2 clocks a field ahead of hardware (774 by field 378); the attract demo seeds its RNG from the H counter after three field-long DMAs and read $F6 for $F8: another demo from NMI 11986	RNG seeded from $2137/$213C; anything that counts on the field length over minutes
Late NMI	the reset code writes $4200 = $81 at V=226; the NMI waited a field, so the NMI count was one behind from frame 5 (6 times in the first 4 minutes, at scene loads)	enabling NMI after a forced-blank load, inside vblank
NMI recognition	the idle loop CMP $0100 / BEQ was interrupted at the other instruction; the 52-clock phase carried into the next fields	every main loop that waits for the NMI
Interrupt sequence	every NMI handler ran 55-60 clocks early	every NMI and IRQ handler; where a raster split lands
$4212 at the access	the NMI handler's BIT $4212 / BNE auto-read wait ended up to 9 passes early; LDA $4212 / BMI left a pass late in the last 20 clocks before line 0	the vblank and auto-joypad waits in nearly every game
Refresh at the beam	the sound upload's CMP $2140 / BNE polled one pass early about every 25 handshakes	the standard SPC700 upload handshake
Reset clock	the first instruction ran 186 clocks early, so every clock of the boot was off by that	anything timed from power-on

The cost of not having them: a host that wants hardware timing has to rebuild them outside the framework, as we did (several
hundred lines that moved the beam after the fact), and that is fragile: walking the beam ahead and syncing back quietly moved
the field grid by 2 clocks, which our own alignment hid until this work removed it. Only Evermore is measured; for other games
the right column says where to look, not what we tested.

Evidence

• Before: Secret of Evermore's host carried all five rules itself (moving the beam after each field, walking it 20 clocks
  ahead at vblank edges, topping up refresh). With that code removed and any one switch off, replay 01 (power-on, attract demo)
  parts from MesenCE 2.2.1:

  SNESRECOMP_SHORT_LINE=0       differs at NMI 60
  SNESRECOMP_RESET_CLOCK=0      differs at NMI 60
  SNESRECOMP_HVBJOY_EXACT=0     differs at NMI 180
  SNESRECOMP_REFRESH_AT_BEAM=0  differs at NMI 60
  

  After: host timing code removed, all switches at their defaults, WRAM, VRAM, CGRAM and OAM compared at every NMI from
  power-on:

  replay 01 (attract demo)       in sync, 8,592 of 8,592 NMIs (native and all-interpreted)
  replay 02 (new game, 6.5 min)  in sync, 23,756 of 23,756 NMIs
  

  Measured together with Return what the PPU fetches for OAM and CGRAM reads during rendering #278 (OAM/CGRAM reads during rendering), which replay 01 needs from NMI 5998.
• Tests, no ROM:

  tests/dma/short_line_test.c           field lengths from reset 306,900 / 357,368 / 357,364 ..., interlace (on and off)
  tests/dma/hvbjoy_refresh_test.c       $4212 vblank and auto-read edges at the access, refresh points
  tests/dma/refresh_charge_test.c       one stall per point crossed, none for a teleport (on and off)
  tests/dma/last_cycle_test.c           last bus cycle: internal, I/O, joypad, WRAM, branches, stack
  tests/frame_driver/exact_rules_test.c scripted CPU on the real beam: NMI position, 60-clock sequence, late NMI
  

  pytest tests/v2 tests/test_snes_cycles.py tests/test_beam_events.py tests/test_beam_frame_state.py tests/snapshot_guard:
  645 passed (macOS). Same speed: 322 frames/s, 94.06 % native (318, 94.05 % before).

Merge Danger

Door: two-way

Every rule has a switch, and the default turns them on only with the master-clock APU timeline. One part does not revert by
switch: savestate v12 adds Snes.fieldParity, so a v12 state does not load in an older build (v11 states load here with a
fresh reset's parity).

Blast Radius: opt-in

Only hosts that call RtlApuUseMasterClock. For them, field lengths, $4212 and refresh timing change, and the frame driver
takes its exact path. Conflicts with #278 in snes_beam_at (a few lines); the resolution is ready, whichever merges second.
Checked on one game and on macOS only.