xtc/docs/PROJECT_FORMAT.md

# Project format v11 (reads v1–v10)

Phases 5–31 use a bounded text format implemented in C99 without a parser library.
The extension is `.xtc`. File operations run exclusively on the control thread.
The project owns its graph, tracks and silent/audio arrangement regions. Audio samples remain in external WAV
files; MIDI notes are stored inline.

## Schema

Records occur in this exact order, separated by LF, including the final record.
Fields use ASCII spaces. Integers are unsigned decimal, without signs.
Counts determine the number of subsequent records:

```text
XTC_PROJECT 11
APP 0 1 0
TEMPO 120000
IDS 1 1
GRAPH 1 2 2 1
NODE 0 1 3e800000
NODE 1 4 00000000
EDGE 0 1
TRACKS 1
TRACK 0 547261636b
MIX 3f800000 00000000 0 0
ROUTE 0 18446744073709551615
CLIPS 1
CLIP 0 0 3840 2880 960 3840 526567696f6e
ASSET -
MIDI 0
AUDIO 3f800000 0 0
NOTES 0 0
AUTOMATION 0
CONTROL_MAPS 16
MAP 0 7 10 64
MAP 1 7 10 64
MAP 2 7 10 64
MAP 3 7 10 64
MAP 4 7 10 64
MAP 5 7 10 64
MAP 6 7 10 64
MAP 7 7 10 64
MAP 8 7 10 64
MAP 9 7 10 64
MAP 10 7 10 64
MAP 11 7 10 64
MAP 12 7 10 64
MAP 13 7 10 64
MAP 14 7 10 64
MAP 15 7 10 64
END
```

- `APP`: producer major/minor/patch; informational, never executed.
- `TEMPO`: constant milli-BPM (20000..300000, default 120000), fixed 4/4.
- `IDS`: next track ID and next clip ID. IDs are uint64, monotonic and never
  reused; UINT64_MAX is reserved. Track and clip ID namespaces are separate.
- `GRAPH`: output-node policy (0/1), next node ID, node count, edge count.
  Nodes retain stable IDs (limited to the host's size_t range). No pointers or
  compiled state are serialized.
- `NODE`: ID, type (1 gain, 2 oscillator, 3 meter, 4 output, 5 audio source, 6 instrument, 8 LADSPA plugin), and exactly eight
  hexadecimal digits encoding the IEEE binary32 parameter bits, independent
  of locale and host byte order. Nonfinite values are rejected.
- `EDGE`: source and destination node IDs. Port directions are implied
  by node type; endpoints, duplicates and cycles are validated by the graph API.
- `TRACK`: ID and hex-encoded name bytes. In v4/v5 each is immediately followed
  by `MIX gain_bits pan_bits mute solo`: finite binary32 gain 0..4, pan -1..1
  encoded as eight hex digits, mute and solo unsigned 0/1.
- `ROUTE`: since v9, required after MIX: bus flag and destination bus ID;
  UINT64_MAX denotes master. See Phase 24 constraints below.
- `CLIP`: ID, track ID, start tick, length tick, source offset tick, source
  length tick, hex-encoded name bytes. Trimming
  preserves source bounds, so a region can re-extend only within its original
  length. Overlaps are allowed and audio clips mix through the audio source node.
- `ASSET`: required after each CLIP in v2/v3/v4/v5; "-" means no audio asset, otherwise
  hex-encoded absolute path bytes, <=511 bytes after decoding, no embedded NUL
  or ASCII control characters. Each clip directly references a WAV; no sample
  data, pointers or DSP state are embedded. Paths can contain spaces.
- `MIDI`: required after each ASSET in v3/v4/v5; 0 for silent/audio regions, 1 for
  MIDI clips. MIDI clips must have no asset.
- `NOTES`: next stable note ID and global note count, followed by that many
  `NOTE id clip_id start_tick length_tick pitch velocity channel` records.
  Notes reference MIDI clips; start/length are source-relative, positive length
  within source bounds. Pitch is 0..127, velocity 1..127, channel 0..15.
  Note IDs are unique, monotonic uint64 IDs with UINT64_MAX reserved.
- `AUTOMATION`: global point count, after the last NOTE and before END in v4/v5.
  Follow with `POINT track_id parameter tick value_bits step` in v10 (v4–v9
  omit step and migrate to linear). Step is 0 linear or 1 hold for the next segment.
  Parameter 0 is gain (0..4), 1 pan (-1..1); values use finite binary32 hex.
  Track must exist; tick <= MAX_TICK; duplicate track/parameter/tick is invalid.
  Record order is preserved; rendering sorts a separate prepared copy.
- Names decode to nonempty valid UTF-8 of at most 63 bytes; ASCII control
  characters and embedded NULs are forbidden. Hex avoids quoting ambiguity.

Ticks use 960 per quarter note, 3840 per bar. All region endpoints are bounded
by 38400000 ticks (10000 bars). Limits: 64 tracks, 1024 clips, 64 graph nodes,
4096 edges, 4096 notes, 1024 automation points, 4 MiB input, 1535 bytes per record, 12 tokens per record. The reader
rejects invalid references, duplicate IDs, inconsistent next IDs, ranges,
truncated records, raw control/non-ASCII bytes, and trailing data.

Unsupported format versions return UNSUPPORTED. Other malformed data returns
FORMAT. The reader explicitly accepts v1 records without ASSET lines, with
node types 1..4, migrating their clips to empty asset paths. Version 2 accepts
node types 1..5 and ASSET records. Both migrate to MIDI=0 and no notes.
Version 3 accepts instrument nodes and MIDI/NOTES records. Versions 1–3 migrate
all tracks to unity gain, centered pan, mute/solo off and no automation.
Saving always writes v11. Unknown versions are rejected without replacing the destination; older
applications without v11 support cannot read current saves. Migration and maximum-path roundtrips are tested.

## Ownership and failures

Load builds and validates a separate candidate, closes the input, then replaces
the destination. Any failure retains the complete previous project. Saving
validates the project, creates a mode-0600 temporary in the destination
directory, writes, flushes, fsyncs, closes and renames over the destination.
Failures before rename retain the previous destination and remove the temporary.
After rename, the parent directory is fsynced. A directory-sync failure reports
I/O error even though the new complete destination may already be present.
Storage and filesystem behavior still determine power-loss durability. Save replaces symlinks themselves and does not preserve
existing file permissions. POSIX I/O is private to the Linux-first persistence
module; public headers contain no POSIX types.

Saved data reproduces the arrangement, graph, track mix and automation, including
ID allocation state. Undo history is not saved. See MIXER.md for pan/interpolation.
Graph node positions, timeline zoom/scroll/selection, transport position and DSP
oscillator phase are GUI/session state and are not saved. Asset references are
absolute in memory; saves use relative paths for media inside the project folder,
otherwise absolute paths. Media are not copied on Save As. Metadata loading preserves
missing references; audio-plan preparation reports unreadable/invalid assets
and renders them silently. Relinking replaces a selected clip's path without
changing its IDs or arrangement. See AUDIO_FILES.md for recovery and limitations.

## Phase 9 extension and recovery policy

Phase 9 kept format v4. Native processor descriptors are process-local and have no
serialized representation; saving a native graph returns UNSUPPORTED and retains
the existing destination. Unknown file versions/types remain rejected. See
NATIVE_API.md for the extension boundary and RECOVERY.md for write-failure, killed
writer and unfinished-recording behavior. Metadata validation is available via
`xtc-check --validate PROJECT` without a GUI or server; it does not load assets.

## Phase 15: version 5 plugin metadata

Phase 15 introduced `XTC_PROJECT 5`; its reader also supported versions 1–4. All v4 records and
order remain the same. Node type 8 is a LADSPA plugin; native pointer-bearing type
7 is still unsupported in files. Immediately after each type-8 `NODE id 8
00000000` record come:

```text
PLUGIN unique_id descriptor_index mono_inputs control_count hex_label
MODULE hex_absolute_module_path
CONTROL port_index float32_hex_bits
```

There are exactly `control_count` CONTROL records. Label length is 1–63 bytes,
module path 1–511 bytes (absolute), both hex-encoded with no embedded NUL. Index
is <1024, mono_inputs 0/1, controls <=16, control port indices <64 strictly
increasing, and values finite. Unique ID fits uint32. The one output is implicit.
Structural validation never opens libraries. On graph compilation the host checks
identity, layout, input control port IDs and rate-adjusted legal ranges against
the module. Unknown nodes/versions fail without mutating the current project.
Missing modules remain preserved by metadata load/save; GUI installation rejects
them with a repair message. No transient DSP state, pointer or plugin binary is
stored. See PLUGIN_HOSTING.md and decision 0009 for supported processors/limits.

## Phase 16: version 6

Phase 16 introduced v6 and its reader accepted v1–v6. After each plugin's CONTROL records,
`PLUGIN_AUTOMATION count` is followed by exactly count (<=32) records:
`PARAM tick port mode float32_hex_bits`. Mode is 0 linear or 1 step. Duplicate
port/tick pairs, nonfinite values, unknown control ports and ticks beyond the
project limit are rejected. Loaded plugin tempo comes from TEMPO. Range and
discrete-mode compatibility are checked on control-thread compilation. Versions
1–5 migrate with empty plugin lanes. See CONTROL.md for timing/ordering.

## Phase 18: version 7 and project folders

Phase 18 introduced v7; its readers migrated v1–v6. Record ordering and field bounds
are otherwise unchanged from v6. v7 ASSET may contain a relative path, interpreted
against the actual containing directory of the named project file (resolved on
the control thread), never the launch working directory. Opening a symlink to a project uses the
resolved target project directory; Save replaces the named entry as before. After loading, all core
asset references are absolute, including missing media. Media inside the save
folder are written relative; references outside remain absolute. Save As resolves
against its new folder and preserves media identity rather than rebasing it by
accident. Copy/move a project folder together with its referenced media to retain
portability. Saving does not copy assets: import/record into that folder first.
Plugin binaries remain trusted absolute module references, not bundled resources.

Empty assets still use `-`. Paths exclude embedded NUL/control bytes and must fit
511 bytes after resolution; overly long resolved references fail transactionally.
Relative references can name outside the folder using ordinary POSIX semantics;
this is not a sandbox. Symlinks follow filesystem behavior, so portability requires
moving their targets too. Old formats continue to require absolute media paths.
No unknown version/node/record is silently discarded; unsupported versions/types
or malformed/trailing data reject the candidate and retain the current project.
Parsing never opens media/plugins or constructs an audio engine.

Save validates the complete model, writes/fsyncs a same-folder temporary, closes
it, atomically renames it, then fsyncs the containing directory. Failures before
rename retain the old destination. A directory sync failure *after* rename reports
I/O error but the new complete file may already be present: inspect/reopen it;
there is no rollback across that durability boundary. Filesystem/hardware power
loss guarantees still depend on the underlying storage. Recording destinations
have their separate no-clobber policy; ordinary project Save intentionally replaces
the selected document. In-memory models and the existing destination are preserved
on load failure. No automatic unknown-field stripping or media deletion occurs.

`make test-portability test test-plugins test-recovery test-recording` covers checked-in
v1–v10 fixtures (including missing v5/v6 plugins), current exact roundtrips, moved-folder
playback from a different cwd, missing media retention, unknown records, every
truncation and deterministic byte mutations, malformed plugin metadata, interrupted
writers and write-limit failures. ASan/UBSan/static analysis additionally audit the
bounded parser. This is regression/fault-injection coverage, not exhaustive fuzzing
or a claim of power-cut qualification.

## Phase 22: version 8 audio clip settings

Phase 22 introduced v8 and reads v1–v7 with unity clip gain and no explicit
clip fades. After **every** clip's `MIDI` record (including silent and MIDI
clips), v8 requires:

```text
AUDIO gain_float32_hex_bits fade_in_ticks fade_out_ticks
```

Gain is finite, linear and within 0..4. Fade durations are unsigned decimal,
0..source_length_tick; zero disables that explicit fade. MIDI clips require
gain exactly 1 and both fades zero. Invalid or missing AUDIO records reject the
whole candidate. All v7 media-folder rules and atomic-save guarantees remain.
Older applications reject v8 rather than silently stripping clip settings.
Rebuild C/C++ consumers against the extended `xtc_clip_info` public struct;
this development release does not promise an unchanged struct binary layout.

Fades follow the audible trimmed clip boundaries, in current-tempo ticks, with
sample endpoints defined in AUDIO_FILES.md. Duplicating retains settings and
source references; splitting keeps the original ID on the left, gives the right
a fresh ID/source offset, retains outer explicit fades and clears inner ones.
No waveform cache or source audio is serialized. Checked-in v1–v7 migrations,
v8 exact roundtrips, malformed new metadata and byte-mutation tests cover the
format change (`make test test-portability test-audio-editing`).

## Phase 24: version 9 mix routing

Phase 24 introduced v9. Immediately after each TRACK's MIX record it writes
`ROUTE is_bus bus_id`. `is_bus` is 0 or 1; `bus_id` is an existing bus track ID,
or decimal UINT64_MAX (18446744073709551615) for master. Buses must route to
master and cannot contain clips. Invalid destinations, flags and nested routing
are rejected. v1–v8 migrate all tracks to ordinary direct-to-master tracks.
Bus controls and automation use existing MIX and POINT records. Older readers
reject v9 rather than discard routing. The public `xtc_track_info` grew fields;
API consumers must rebuild. Fixtures cover v1–v8 migration and tests exercise
v9 exact roundtrips and malformed routing.

## Phase 26: version 10 interpolation and controller bindings

Phase 26 introduced v10 with migration from v1–v9. Track/bus automation points are
`POINT track_id parameter tick value_bits step`, where step is 0 linear or 1
hold for the segment starting there. v1–v9 points migrate to step=0. LADSPA
`PARAM` encoding remains unchanged (already has interpolation mode).

Immediately after all POINT records, before END, v10 requires `CONTROL_MAPS 16`
and exactly 16 records in ascending channel order:
`MAP channel volume_CC pan_CC sustain_CC`. Channels are 0..15. Controllers are
0..127 or 128 (disabled); enabled numbers must be distinct within a row and may
not be 120 or 123. Duplicate/missing/out-of-order channels, invalid CCs, reserved
bindings and invalid interpolation modes reject the complete candidate. v1–v9
supply default 7/10/64 bindings on every channel. No event history/current live
controller values are serialized. Older readers reject v10.

Fixtures cover v1–v10 loading and migration. Exact roundtrips, malformed metadata, step/linear
renders and remapped live events cover the new writer/reader. Public clients
must rebuild and initialize the appended `xtc_automation_point.step` field.

## Phase 28 compatibility and diagnostic contract

The current reader accepts versions 1–11 and the writer always emits version 11.
Phase 28 made no format change; Phase 31 adds plugin format identity in v11. Saving an older document upgrades it;
retain an original backup if an older application must still open it. APP is
informational; compatibility depends on the format version, not the producer's
application version. A future incompatible schema requires a new format number.
Public C/C++ consumers must rebuild when public structs change.

| Version | Added representation | Defaults when loading earlier versions |
| --- | --- | --- |
| 1 | Graph, tracks, silent regions | Baseline |
| 2 | External WAV references | Empty asset references |
| 3 | MIDI clips and notes | Non-MIDI clips, no notes |
| 4 | Track mix and automation | Unity gain, center pan, mute/solo off, no points |
| 5 | LADSPA metadata | No plugin nodes |
| 6 | Plugin automation | Empty plugin lanes |
| 7 | Relative media references | Earlier media paths must be absolute |
| 8 | Clip gain and fades | Unity gain, zero explicit fades |
| 9 | Bus routing | Ordinary tracks routed directly to master |
| 10 | Track interpolation and controller bindings | Linear points; CC 7/10/64 on every channel |

The grammar has no optional extension namespace. Unknown records, extra fields,
node types or trailing data cannot be preserved by this model and are rejected
rather than silently removed. Supported older records migrate in a separate
candidate and pass complete model validation before adoption. Failed parsing,
reference validation or close retains the existing destination model. GUI Open
also prepares the candidate before replacing the live document; missing or
incompatible plugins retain the current document and identify the module.
Missing media references remain available for relinking and render silently.

`xtc_project_load_diagnostic` adds a copied result, parsed format version (zero
if unavailable), expected record/stage and message. Syntax locations are 1-based;
line zero denotes I/O, size, path resolution or whole-model validation. MODEL
errors do not identify the original line of a cross-reference. The original
`xtc_project_load` remains available with identical transactional semantics.

`xtc-check --validate PROJECT` checks bounded metadata without loading dependencies.
`xtc-check --inspect PROJECT` additionally lists resolved media and module paths,
plugin labels, unique IDs and descriptor indices. It checks readable regular-file
availability only. Exit status is 0 when references are available, 2 when valid
metadata has missing/unreadable references, and 1 for invalid metadata or I/O.
Neither command executes plugin code, decodes WAV data, connects JACK or verifies
hardware. A readable file can still be an invalid WAV or incompatible plugin;
actual preparation performs those checks. Paths and labels are escaped for
terminal output. There is no search-path substitution or automatic plugin relink.

Portability currently targets Linux/POSIX. The text representation is independent
of byte order and locale and requires IEEE binary32 floats. IDs used by the graph
must fit the receiving host's size_t; a 64-bit file is not guaranteed to fit a
32-bit build. Paths follow POSIX semantics; Windows drive letters and separators
are not translated. Relative media can travel with their project folder, while
absolute references and symlink targets must be restored separately. LADSPA
modules are absolute installed references: match architecture, ABI, descriptor
identity, ports and legal controls on the destination. Plugin binaries and state
are not bundled. JACK availability, sample rate, block limits and device routing
are runtime dependencies; loading metadata does not restore device connections.

`make test` includes v1–v10 migration, malformed/truncated/mutated/oversized input,
retained-state diagnostics, CLI dependency inspection, and deterministic save
faults at flush, file sync, close, rename and directory sync. Pre-rename faults
preserve the old bytes and remove the temporary. After rename, a sync failure
reports I/O while leaving a complete new document; inspect it before retrying.
Fault wrappers are linked into tests only. These tests complement killed-writer
coverage in `make test-recovery`; they do not qualify power-cut behavior.

## Phase 31: version 11 plugin format identity

Immediately after every plugin MODULE record, v11 requires:

```text
FORMAT format_number hex_identity_or_dash
```

0 = LADSPA, 1 = CLAP, 2 = VST2, 3 = VST3. Identity is `-` for LADSPA/VST2;
CLAP uses its stable plugin ID (1–127 non-NUL bytes); VST3 uses the 32-character lowercase
hex class ID string, itself hex-encoded like other byte strings. PLUGIN's unique
ID remains the VST2/LADSPA integer ID (zero for CLAP/VST3), descriptor index and
label. MODULE is still the actual absolute binary path. Versions 1–10 migrate
to format 0 and empty extra identity. Older readers reject v11. The public plugin
struct has grown; C/C++ consumers must rebuild and initialize it to zero.

Modern adapters accept one-input mono effects, default configuration only, with
zero controls and automation points. Unsupported modern control metadata is
rejected, not discarded. State/editor/preset changes are not exposed, so no opaque
state is serialized. Reloading recreates plugin defaults, which can change after
a plugin update or with its external configuration. Matching ID/name/layout is
not a binary hash or sound-reproducibility guarantee. See PLUGIN_HOSTING.md.

Metadata validation and inspection never initialize modules. Actual graph
preparation verifies format, descriptor, stable identity, name and layout before
publishing. Missing module metadata survives C load/save; GUI preparation retains
the existing document on failure. No automatic path search or plugin substitution
occurs. Current and historical migration fixtures, malformed fields, identity
mismatch, missing-module and exact-render roundtrips cover this extension.