Parse SF2 modulator chunks (pmod/imod) and store on SF2Zone. Full modulator
evaluation per SF2 spec 9.5.3: source extraction (velocity, key, CC, pitch
wheel), transform curves (concave/convex/switch, unipolar/bipolar), two-source
multiplication, absolute value transform. All 10 default modulators from SF2
spec 8.4.1-8.4.10 with zone override logic. Replaces hardcoded velocity curve.

Parsing:
- SF2Modulator struct, parse_modulators for pmod/imod chunks (10 bytes/entry)
- Source unpacking: sf2_mod_src_index/cc/negative/bipolar/type
- Modulators stored on SF2Zone alongside generators

Evaluation:
- sf2_concave/sf2_convex: SF2 spec page 73 transform curves
- sf2_mod_source_value: velocity, key, CC, pitch wheel, pitch wheel sensitivity
- sf2_mod_transform_source: unipolar/bipolar, positive/negative, all 4 map types
- sf2_mod_get_value: amount * transform(src1) * transform(src2), abs transform
- sf2_apply_modulators: accumulate per-generator offsets from zone modulators

Default modulators:
- SF2_DEFAULT_MODS: 10 defaults (vel->atten, vel->filter, pressure->vib, CC1->vib,
  CC7->atten, CC10->pan, CC11->atten, CC91->reverb, CC93->chorus, pitchwheel->tune)
- sf2_mod_identity_match + sf2_apply_default_modulators with override logic
- Removed hardcoded sf2_velocity_attenuation (now handled by default mod #1)

Application in sf2_create_c_voice:
- Compute gen_offsets[64] from zone + global + preset modulators + defaults
- Apply offsets to pitch, filter FC/Q, attenuation, pan, all LFO/env routing

Also: midi_init_reverb() public function for offline rendering with reverb

20 new tests: source unpacking, evaluation (6 cases), parse counts (FluidR3=746,
TimGM6mb=455), voice velocity dynamics, default mod dynamics, override logic

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Fixes 7 review comments in one push:

* MCP stdio safety (#1, #5): read_import and compile_and_collect both
  print warnings/failures to stdout, which corrupts the JSON-RPC
  protocol stream when called from the MCP server. Add `quiet : bool
  = false` to both helpers (default preserves CLI behavior); MCP
  find_duplicates passes true and surfaces failure counts via the
  returned envelope.

* Test portability (#2, #3): replace hard-coded /tmp/... corpus path in
  test_find_duplicates_corpus_not_found with create_temp_file_result +
  immediate remove() — guaranteed-nonexistent path that works on
  Windows runners.

* Newline-delimited paths (#4): MCP do_find_duplicates now normalizes
  \n to , before parse_file_list, matching the README's
  documented "git diff --name-only | …" pipeline use case.

* Project (.das_project) plumbing (#6): add `project` to the
  find_duplicates schema, dispatch, and handler signature; thread
  through to compile_and_collect, which now takes
  `project : string = ""` and uses make_file_access(project) instead of
  the previous hardcoded "" — matching the convention used by
  compile_check / lint / find_symbol.

* Deterministic compile order (#7): both the MCP tool and CLI main.das
  build a sorted array<string> from `keys(against_files)` and iterate
  that, instead of the unspecified-order table-key iteration. Makes
  per-candidate report ordering and --check exit behavior reproducible
  across runs.

Drive-by: PERF006 fix in resolve_against_files (reserve before push
loop).

All test suites still green: 6987 in tests/, 54 find_dupes, 181 MCP
(includes 5 updated find_duplicates tests for new signature). MCP
end-to-end smoke confirms newline-delimited paths and project
parameter both reach the handler.

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

PR GaijinEntertainment#2575 review round 3 — three intertwined Option<T> bugs surfaced by Copilot.

1. Generic `sql_bind($Option<auto(TT)>)` fallback in sqlite_boost mirrors
   the existing `sql_storage_enum_for(type<$Option<auto(TT)>>)` pattern.
   Uses `_::sql_bind(default<TT>)` so caller-module adapters resolve;
   without `_::`, recursion runs in sqlite_boost's module and misses
   `@sql_json`/`@sql_blob`-generated overloads.

2. `emit_json_blob_adapters_for_struct` now unwraps Option<T> before
   adapter codegen. Runtime `sql_bind_to_stmt(Option<T>)` strips the
   Option BEFORE calling user `sql_bind`; registering the adapter for
   `Option<NoteData>` was dead code — actual call lands on `sql_bind(NoteData)`.
   Adapter dedup now naturally collapses bare T and Option<T> in the
   same struct (both key as T).

3. Drop the macro-time Option unwrap in `apply_schema_from` from the
   prior round; the witness is back to `_::sql_bind(default<field_type>)`,
   with rail #1 handling the unwrap uniformly.

Tests:
- test_96_schema_from_partial_body: @sql_json + Option, @sql_blob + Option
  via schema_from path (contract + roundtrip).
- test_64_json_columns: standalone @sql_json + Option, plus dedup edge
  case (bare X + Option<X> in same struct).
- test_65_blob_struct: standalone @sql_blob + Option.

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