Release Inseglet v1.13.2alpha

[jlivingston-Cipher]

Changelog

All notable changes to Inseglet are documented here. The format is based on
Keep a Changelog, and the project aims to follow
Semantic Versioning.

[Unreleased]

[1.13.2] — 2026-08-24

Documentation release. No functional change; the only source edit is a comment.

The standards basis now says which loudness engine implements which revision of ITU-R BS.1770.
Two paths ship and they cite different revisions on purpose: REAPER's native RENDER_STATS
program measure — analysis.meter's program loudness and every deliverable-spec check — follows
BS.1770-5, while the in-box C++ gated engine, which measures stems, objects, the isolated
dialog bus and folds that exist only in memory, implements BS.1770-4 and is pinned to it by
unit test. docs/CONVENTIONS.md had said "throughout", which was the only false statement.

README.md claimed CI builds the three-OS matrix on every push. It does not: CI compiles and
unit-tests on Ubuntu on every push, and the macOS / Windows / Linux matrix runs on a version tag
or a manual dispatch.

docs/MANUAL.md hedged on the published ReaPack index URL. It now states it, and states the
macOS signing posture plainly: the shipped reaper_mcp.dylib is ad-hoc codesigned — the
signature Apple Silicon requires to load it at all — and is not notarized.

Surface 190 tools / 4 resources / 5 prompts — unmoved. Unit suite 28, unmoved.

[1.13.1] — 2026-08-18

Hygiene release: a user-facing warning no longer cites an internal engineering document.

analysis.meter's CROSS_NO_SIBLING warning — the one that fires when a digital-silence
render has no independent read path to check it against — carried a parenthetical citation
to an internal document. The sentence is unchanged in meaning; the citation is removed.

Surface 190 tools / 4 resources / 5 prompts — unmoved. Unit suite 28, unmoved.

[1.13.0] — 2026-08-18

Refusal release: analysis.meter will no longer REPORT a digital silence that two
independent read paths disagree about over the same bounded window.

analysis.meter renders to measure. When that render comes back all-zero the tool used to
report the silence as a measurement, and a dead render path is byte-for-byte
indistinguishable from a genuinely silent one -- the same -144.0 on every channel, the
same rawStats carrying LENGTH alone, no error key, and elapsed times that do not
discriminate either. The tool now cross-reads the same window through
analysis.accessor_meter, which reads the track without rendering, and when the two paths
disagree it refuses the CLAIM rather than the call: render_silent_unconfirmed. Pass
allowSilent and it reports anyway, attaching BOTH readings and windowMatched so the
disagreement is in the payload rather than in a footnote.

Two things it deliberately does NOT do. It does not refuse when the two reads covered
DIFFERENT windows -- an unmatched window downgrades the refusal to a report and says so,
because two readings over different windows are not a comparison. And it does not treat
"the accessor sees content" as proof the render is dead: the accessor reads PRE track FX,
volume and pan, so a MUTED track legitimately reads content and renders to silence.

The warning now names WHICH branch produced the cross-read, as crossRead.pathKind:
accessor_read, accessor_failed, or no_sibling_path. It previously selected on a bare
availability boolean, so a read whose accessor path EXISTED AND FAILED was reported as
"no independent read path exists for this target" -- the right verdict attached to the
wrong reason. A fail-closed fourth branch names an unaccounted state rather than borrowing
the nearest plausible one.

*** Behaviour change: a call that previously returned a silent measurement can now refuse.
*** Pass allowSilent:true for the previous behaviour, with both readings attached.

Surface 190 tools / 4 resources / 5 prompts -- UNMOVED. ctest 28, unmoved.

[1.12.0] — 2026-08-15

Inseglet can now tell you whether a take's media is actually there. The surface grows to
190 tools / 4 resources / 5 prompts; the release is purely additive — no tool changes
what it returns, refuses anything it previously accepted, or reads different samples.

Added

• reaper://track/{trackIndex}/item/{itemIndex}/take/{takeIndex}/source — a source-validity
  panel. A per-source readout for one take: the ReaScript-API rows (file name, type, channel
  count, sample rate, length) alongside the SDK vtable rows, including
  PCM_source::IsAvailable(). IsAvailable is not exposed by the ReaScript API at any
  version, so this is the only route to a take's source-validity state — online, offline, or
  missing.
  ⚠️ Read vtable.controlPassed before trusting any vtable row. When it is false, the SDK
  header and the running REAPER disagree about the vtable layout and no vtable figure is
  believable. The panel carries that control itself rather than letting the numbers quietly lie.
• A sourceValidity block on every accessor read — analysis.accessor_meter,
  analysis.accessor_levels and analysis.accessor_gaps. It reports how many sources the read
  interrogated, how many reported unavailable, the raw sample rate / channel count / length
  the source gave, and defaulted, which names every field a fallback stood in for. When a
  source is unavailable the block carries the remedy: action.run {command: 40101} revalidates
  the project's sources.
  available is null, never true, when a read queried no source at all, and the three
  single-source raw figures are null rather than 0 on any read that did not interrogate
  exactly one source — a zero there would be indistinguishable from a source reporting zero.

Fixed

• The accessor reported an identical read whether a take's media was there or not.
  src/audio_accessor.h substituted a plausible sample rate for a source that reported 0, and
  a channel count of 2 for one that reported less than 1 — exactly the two fields REAPER
  zeroes when it cannot open a source. The evidence that anything was missing was destroyed
  before it could be read. The reads still fall back, so behaviour is unchanged, but every
  substitution now names itself and the raw readings travel beside it.
• The track-read path interrogated no source at all, so removing the fallbacks alone would
  not have helped it. It now asks every take of every item on the track.
• docs/REFERENCE.md regenerated — it still advertised three resources.

Testing

• New unit.source_probe (unit suite 27 → 28). The predicate lives in its own REAPER-free
  header so its whole truth table runs on a host build, and it carries two controls that must
  fire: the pre-fix defaults reproducing the identical-read result, and the predicate
  stripped of its quarter-note guard reporting an ordinary MIDI take as a broken source — a
  source measured in quarter notes legitimately has no sample rate, and a detector that fires on
  healthy content is a detector its users switch off.

[1.11.0] — 2026-08-12

The export tools can now fail to measure. The surface is unchanged at 190 tools /
3 resources / 5 prompts, but three tools gain one parameter — a schema change the
tool-count triple cannot see.

Fixed

• spatial.export_adm, spatial.export_damf and spatial.export_loom_manifest authored
  deliverables from renders that never happened. All three read a render in which every
  sample of every channel was exactly zero and emitted a complete, well-formed,
  permanently archivable file whose loudness claims were the arithmetic image of "no
  samples": -120 dB RMS, a -70 LUFS floor, 0.0 active fraction. The previous
  render_empty guard only ever caught a render of zero length. All three now refuse
  with a named render_silent error and write nothing.
  ⚠️ BREAKING: authoring a deliberately silent master now requires allowSilent: true,
  which reproduces the previous bytes exactly — verified byte-identical on three fixtures.
• The guard is the whole image, never a threshold. A legitimate 7.1.4 deliverable with a
  single channel fed has eleven of twelve channels on the exact floor triple, so a
  per-channel rule would reject real work. One 1e-9 sample clears it: whether a quiet
  master is too quiet is not this tool's judgement to make, and the -70 floor is correct.
• It was never one site. export_damf authored its triad from the same unmeasured
  render, and export_loom_manifest emitted the same floor triple through
  intentFillLevels / intentBedLufs plus a whole-bed expectLufs: -70.

Documentation

• docs/REFERENCE.md regenerated. allowSilent had existed in three tools' input
  schemas and in no user-facing document, because the standing surface gate counts tools,
  resources and prompts — a new parameter is invisible to it.

[1.10.0] — 2026-08-11

A correctness release. The surface is unchanged at 190 tools / 3 resources / 5 prompts —
no new verbs, no new parameters. Released as a minor rather than a patch because reported
values move.

⚠️ Absolute loudness figures are not numerically comparable with v1.9.0 and earlier.
Figures reported by any prior release are low by 0.047099 / 0.043277 / 0.023566 /
0.021652 LU at 44.1 / 48 / 88.2 / 96 kHz. Relative quantities — channel balances,
deltas, ratios, LRA, activity fractions — are unaffected: the correction is a pure gain at
every frequency, so the shift cancels. To compare archived absolute figures against v1.10.0
output, add the constant for your sample rate.

Fixed

• The BS.1770-4 RLB filter was non-conformant, and every per-channel loudness figure read
  low by a constant. The K-weighting stage-2 highpass divided its numerator by a0, forcing
  the passband gain to exactly 1.000000000000. BS.1770-4 Table 1 keeps that numerator as
  exactly [1, -2, 1] — it is not normalised — and its gain is a0 = 1.004994898715 at
  48 kHz. The -0.691 calibration term is defined against the tabulated filter, so the
  normalised form read low at every frequency. Moved by the fix: analysis.meter's
  channelsDetail[].kLevelLkfs, accessor_meter, object_loudness, stem_loudness,
  dialog_loudness, downmix_loudness, gatedLoudness(), and the intent sidecar's
  expectLufs claims.
• The K-weighting shelf gain constant was truncated to 12 significant digits where the
  reference carries 16, reaching b1 as a 1.044942e-12 error against Table 1. Worth about
  3.3e-12 LU — below every measurement — but it accounted for the whole of a discrepancy
  that had been attributed to the table's own precision rather than to this code.

Verified

• Live against a running REAPER on 28 windows: the shift reproduced to 3.5e-12 LU, and the
  cross-engine residual column came back invariant to 3.5e-12 LU.
• Cross-checked against a second, independent BS.1770-4 implementation on 13 rows: the two
  now agree to 1.7799e-12 LU, where before the fix they differed by exactly the constant
  above — with this implementation on the low side and the other one conformant.
• unit.meter now gates all five coefficients of both filter stages against Table 1 at
  1e-12, with a negative control that rejects the pre-fix numerator. The previous check
  gated the K-curve against flatness at 1 kHz to ±1.2 dB, which is 28× too loose to see a
  0.043 dB error.

[1.9.0] — 2026-08-11

A correctness release. No new tools: the surface is unchanged at 190 tools / 3 resources /
5 prompts, and this release adds one parameter and no verbs.

Fixed

• analysis.verify_routing reported false faults on 22.2 beds. 22.2 is the only bed layout
  that declares two LFE slots, and the identifier-tone plan assigned the same 40 Hz LFE tone to
  both — so the tone set was not unique and the detector could not tell a real fault from a
  collision. A correctly routed 24-channel render came back ok=false, with channel 3
  reported as bleed at a margin of exactly 0.000000000 dB and channel 9 as duplicated, both
  at full plan coverage — there was nothing in the reported numbers to indicate the verdict was
  wrong. Reaching it required following the tool's own guidance and passing back the
  lfeChannels it echoes; callers who omitted them saw a clean 24/24, which is why it went
  unnoticed. The detector now refuses a plan whose tones are not unique and returns no
  per-channel verdicts at all, rather than attaching a warning to an answer it cannot stand
  behind. Only 22.2 was affected; every other layout declares exactly one LFE.
• 7.1.2 beds were measured without BS.1770-4 channel weights, reading 0.350 LU low. Channel
  labelling enumerated the other bed widths but had no entry for 10 channels, and the weight
  table derives entirely from labels — so a 7.1.2 bed silently received a flat weight vector
  while being reported as "multichannel". 7.1.2 is the default bed layout in three spatial
  tools, and the predicted bed loudness written into the intent sidecar inherited the error.
  Every accepted layout now comes from one canonical table that the test suite walks, so a
  layout added later cannot repeat it.
• spatial.inject_identity_tones advertised "7.1.2" and then refused it at runtime. The
  layout appeared in the tool's own schema enum while its width lookup had no entry for it. All
  bed-layout widths now derive from the single canonical table, and five separate hand-written
  width mappings have been removed. A new guard walks the whole tool registry and asserts that
  every layout any tool advertises is actually accepted, so this class cannot return.
• spatial.inject_identity_tones could serve stale audio after a parameter change. Generated
  files were named by slot index alone, so two calls with different parameters wrote to the same
  path. REAPER caches PCM per path from the second access onward, after which a rewrite of that
  path is not picked up — the tool reported the parameters you asked for while REAPER played the
  earlier audio. Filenames now include a hash of the audio content, so identical parameters still
  produce identical paths (determinism is unchanged, and the cache hit is then correct) while
  different parameters can no longer collide.

Added

• spatial.inject_identity_tones gains constantHz — put a single frequency on every
  channel, LFE included, instead of the per-channel identifier ladder. Useful for level and
  loudness measurement, where a common tone across channels is what you want. The response now
  reports routingDetectable, and no longer suggests routing verification for a plan whose
  tones are deliberately identical, since routing cannot be inferred from such a render.

Changed

• Documentation and packaging metadata corrected: the citation file and the manual's version
  header had drifted behind the released version.

[1.8.0] — 2026-08-06

Feature release: IAMF ingest — an Inseglet ADM export is now accepted by
iamf-tools — plus in-DAW channel-identity QC and an
FX-parameter authoring seam. Surface 190 tools / 3 resources / 5 prompts (was 187 / 3 / 4);
unit suite 23 → 27.

Added

• spatial.export_adm gains dolbyMetadataChunk — the opt-in dbmd chunk that lets an
  export's objects survive IAMF ingest. "none" (the default) is byte-identical to the
  previous writer; "placeholder" emits a 32-byte dbmd after fmt and before chna and
  renames the bed to Dolby's RoomCentric* vocabulary — one switch, because the consumer makes
  them one: iamf-tools reads the chunk's existence only, that existence selects its Dolby
  validation path, and only that path consults channel names at all. Without a dbmd the importer
  takes its default path, rejects every audioObject as "Not under common definition" and fails
  the encode with "No audioObject present" — no IAMF file is produced at all, not even a bed-only
  one (measured 2026-08-06 against iamf-tools main 19019e3; certified Dolby/EBU renderers are
  unaffected and ingest either form). Independent of profile — profile:"dolby-atmos" still
  emits no dbmd. Fails closed, with the remedy named, on every combination that consumer
  rejects: bit depth ≠ 24 and sample rate ∉ {48 kHz, 96 kHz} (both refused at the reader, before
  any metadata is parsed, so taking the switch unguarded would be worse than declining it), and any
  bed outside 5.1 / 7.1 / 7.1.2, because the importer's pack-layout check is an allow-list of
  eight ordered layout strings — a legal channel-name set is necessary but not sufficient. Taking
  the switch asserts the file is a Dolby ADM deliverable, so a runtime advisory discloses the
  trade on four surfaces — spatial.export_adm, analysis.adm_inspect,
  analysis.adm_profile_check (in warnings, deliberately not violations: the Atmos profile
  does not require a dbmd) and a new fifth prompt author_dolby_adm — firing only when objects
  are present and the switch is off, because a hint that always fires gets filtered. Prompts
  4 → 5; tool surface unchanged at 190 — a parameter is not a verb. ⚠️ The advisory is a claim
  about third-party software: it carries its measurement date, because upstream is actively
  reworking this importer.

• envelope.ensure_fx_envelope — the FX-parameter authoring seam. envelope.add_point could
  only write to an already-active envelope, and autoActivate reaches only REAPER's built-in
  track envelopes; FX-parameter envelopes need the FX + parameter context. This verb ensures the
  envelope exists and returns the name add_point looks up. Address the parameter by index or
  by name — resolution is tiered (exact → case-insensitive → substring) and an ambiguous request
  is refused and names its candidates rather than picking one, so paramName:"Angle" on an IEM
  encoder reports all three Angle parameters instead of silently automating azimuth. Reports
  addPointSafe: when two FX on a track expose an identically-named parameter the envelope name
  does not identify one envelope. Idempotent. Surface 189 → 190 tools.

• In-DAW channel-identity QC — spatial.inject_identity_tones + analysis.verify_routing.
  The first lays a unique identifier sine per bed channel or object track (313 + 139·k Hz, LFE
  40 Hz, −18 dBFS); the second reads any downstream point — an external or rendered WAV, or a
  render-free track accessor — and reports which tone actually arrived on each channel: identity /
  swapped / duplicated / dropped / bleed / silent, each with a measured marginDb and
  planCoverage. Catches swapped, dropped and duplicated channels and inter-channel bleed inside
  the session, before export, and independently verifies the send encodings. Surface
  187 → 189 tools (spatial 23 → 24, analysis 18 → 19).

Fixed

• ADM audioObject emission order — spatial.export_adm output is no longer refused by
  iamf-tools' ADM importer. The writer emitted each object as audioObject → audioPackFormat →
  audioChannelFormat (whose blocks carry <gain>) → next audioObject. That importer sets its
  gain tag for any element named gain without checking the parent, and registers no
  end-element handler, so the tag latches and parent is never reset; an <audioObject> starting
  while the tag is live makes it parse our indentation as a gain and refuse the whole file with
  INVALID_ARGUMENT: Failed to parse gain. The failure is a conjunction of element order and
  inter-element whitespace — neither alone triggers it — and is invariant to the gain value,
  position format, coordinate mode and dbmd, which is why it resisted diagnosis. Every
  audioObject is now emitted before any audioChannelFormat. No emitted byte changed — only
  the order in which the three groups are joined; ids, IDRefs and audioTrackUID numbering are
  untouched.

Changed

• Live-gate surface checks assert a floor, not an equality. Each scripts/verify_*.py gate
  pinned the absolute tool count as of the beat that wrote it, so every earlier gate failed its
  first check against any later build. The count only ever supported a floor — the gate's real
  claim, that its own verbs are on the wire, is asserted by name on the following lines — and a
  floor still catches the failure these gates exist to catch: running one against an older
  installed binary than it needs.

[1.7.0] — 2026-08-05

Feature release: the authoring-side intent sidecar + representation-aware positioning.
No surface-count change (187 tools / 3 resources / 4 prompts).

Added

• Intent sidecar emission — intentSidecar: true on spatial.export_adm and
  spatial.export_loom_manifest writes the session's own predictions beside the export
  (<base>.intent.json / manifest.intent.json, schema intent: 0): a bed roster with
  measured per-channel expect* levels (the whole-bed figure under the BS.1770-4-conformant
  weights), object trajectories at block reach times, SN3D projection-decode
  dominance/coverage predictions (order 3), and per-ACN scene RMS. Never a declared
  deliverable loudness — every level is an expect* claim to be checked, not copied; the
  consumer is iamf-sentinel-pro's
  sentinel intent-compare (findings S-340…S-346). Beds outside the consumer's judged set
  (2.0/5.1/7.1/7.1.4) and v0 VO stems emit no claims (absent block = absent claim). Emitted
  expectLufs claims floor at −70 LKFS (the BS.1770 absolute-gate silence convention) so
  silent bed channels round-trip clean while added content is still caught — verified end to
  end through a live REAPER → export → consumer gate. Sidecar write failure is a warning,
  never a failed export. New SDK-free src/intent_sidecar.h, pinned by the
  unit.intent_sidecar suite (byte-goldens held cross-platform).
• One in-source version constant — src/inseglet_version.h now feeds
  serverInfo.version, the discovery file, and the sidecar's producer stamp, so the
  version-drift class (a hardcoded string nobody bumped) can no longer recur one literal at
  a time.

Fixed

• spatial.set_source_position drives the representation the exporter reads —
  angle-representation panners (IEM/SPARTA-style encoders exposing Azimuth/Elevation
  params) are now driven in azimuth/elevation, derived from a cartesian request when needed
  via each param's own degree range; partial writes preserve the untouched axis (z alone
  never recenters azimuth; a single horizontal component composes against the current
  azimuth). Previously the bare-letter name scan could land on Quaternion X for x — and
  on aZimuth Angle for z, so a height write silently rewrote azimuth — moving params
  spatial.export_adm's trajectory sampling never reads. The cartesian path
  (ReaSurroundPan) is unchanged (regression-pinned against the old scan) and now refuses
  quaternion params with an explanatory message; auto-detect finds angle panners without an
  explicit fx index; explicit paramX/paramY/paramZ or normalizedInput still forces
  deterministic cartesian writes. New SDK-free src/position_params.h +
  unit.position_params (the negative control reproduces the old defect in-test).

[1.6.0] — 2026-08-04

Feature release: IAMF delivery. Surface 186 → 187 tools; expert prompts 3 → 4.

Added

• spatial.export_loom_manifest (Render) — a one-call bridge to the open
  iamf-loom packager: renders the
  session's bed (stereo/5.1/7.1.4) and/or ambisonic scene (order 1–4, ACN/SN3D) plus
  per-language VO stems as 48 kHz integer-PCM WAVs in Loom's channel order, and emits the
  loom: 0 manifest (+ optional season.yaml) beside them — loom compile accepts the
  output as-is, and the response echoes the exact next command. The channel mapping is
  identity (label spellings only: Lsr/Rsr==Lrs/Rrs, Ltr/Rtr==Ltb/Rtb), verified with
  per-channel identity tones. Never emits a loudness value (Loom measures — that is the
  point of it); objects fail closed toward the ADM route (spatial.export_adm);
  known-invalid target combinations fail closed (youtube-without-video, archive/normalize
  constraints, non-48k renders).
• deliver_to_iamf expert prompt (prompts 3 → 4) — inventory → per-mix
  channel-budget check (a 7.1.4 bed + an order-3 scene + a stereo VO = 30 channels exceeds
  IAMF base_enhanced's 28-per-mix cap; order 2 fits at 23) → dryRun preview → export →
  loom compile / loom run with the iamf-sentinel gate.
• MANUAL §9 “Delivering to IAMF” — the pipeline end to end, the channel-budget
  note, and the validate/repair loop; REFERENCE.md regenerated (187 tools / 3 resources /
  4 prompts).

[1.5.1] — 2026-08-04

Correctness-first patch release; no surface change (186 tools).

Fixed

• BS.1770-4 bed channel weights are now read from Tables 4/5 exactly (boundary-inclusive:
  the 1.41 surround weight applies only at ear level within |az| 60°–120°). Two cells move:
  rear surrounds Lsr/Rsr (M±135) 1.41 → 1.00, and wides Lw/Rw (M±060, 9.1.6 — exactly on
  the inclusive boundary) 1.00 → 1.41. Every loudness tool inherits (meter, stem / dialog /
  object loudness, loudness timelines, deliverable checks). Readings move only where the
  affected channels carry energy: rear-bearing 7.1-family content reads cooler (equal-energy
  7.1.4 ≈ 0.3 LU), 9.1.6 wide-bearing content warmer; 5.1 / stereo / 22.2 are unchanged
  (22.2 stays documented-unweighted). The weight table now lives in an SDK-free
  src/bed_weights.h, pinned by the new unit.bed_weights suite whose expected vectors
  mirror iamf-sentinel-pro's independently published BS.1770-4 conformance test; a negative
  control fails at exactly the four moved cells, and the power-domain delta is
  1.4922 dB = 10·log₁₀(1.41).
• ADM writer: object azimuths wrap to (−180°, 180°] and elevations clamp to ±90° at
  spherical block serialization (an az = 270° trajectory no longer produces a file the EBU
  EAR renderer hard-rejects); the LFE <frequency> element is emitted at its BS.2076
  position as an audioChannelFormat child; BW64 output carries the real ds64.riffSize
  instead of 0.

Changed

• analysis.adm_profile_check classifies pack types from chna packRef IDs, so
  common-definitions ADM is no longer invisible: HOA / binaural / Matrix content, beds beyond
  the profile's list, multi-programme and non-24-bit files now correctly report
  non-conformant (previously they blind-passed). Bed layouts gate against the published
  BS.2094 common-definitions table plus a channel-count check, and the interpolation check
  verifies the actual per-block ramp (new violation code interpolation_ramp, distinct from
  presence). Verdicts CHANGE on common-definitions ADM — third-party masters that
  previously "passed" may now correctly fail.
• analysis.adm_inspect reports a bounded per-channel summary (adm.channels[]), a
  packTypes census (with a common-definitions flag), and a corrected ADM version readout.

Fixed

• Version reporting: the MCP initialize handshake now reports serverInfo.version as the package
  version; it had been hardcoded to an early 0.1.0 and was out of sync with the discovery file and the
  ReaPack package metadata (the index and package header are likewise brought to the current version).
• Windows build: the cross-platform release build failed to compile on MSVC with
  error C2065: 'M_PI' undeclared in the spatial tools. MSVC's <cmath> only defines M_PI when
  _USE_MATH_DEFINES is set before the header is included — glibc and Apple's libc++ expose it
  unconditionally, so the Linux and macOS builds were unaffected. Fixed by defining _USE_MATH_DEFINES
  project-wide for MSVC in CMakeLists.txt, and added an explicit ws2_32 link on Windows so non-MSVC
  toolchains (MinGW / clang) link the loopback HTTP sockets too.
• Windows tests: the metering WAV round-trip test wrote scratch files to a hardcoded /tmp/ path, which
  doesn't exist on Windows — the failed write left an empty buffer that was then read out of bounds and
  crashed the test. It now uses the OS temp directory via std::filesystem::temp_directory_path().

[1.5.0] — 2026-07-10

Added

• analysis.object_decode_timeline — a read-only, purely geometric decode-coverage timeline for
  immersive objects, complementing analysis.decode_coverage (which decodes a whole ambisonic scene
  bus). For each object it reads the position trajectory — sampled from a live object track's panner
  automation (azimuth/elevation/distance) or supplied as explicit keyframes — and at every time slice
  encodes the object direction to SN3D spherical harmonics and projection-decodes it onto the chosen
  delivery layout's nominal loudspeakers, picking the dominant speaker. It reports the dominant-speaker
  timeline, the migration run-path, per-speaker time-weighted dwell fractions, the great-circle angular
  travel and elevation reach, and a time-weighted coverage footprint (hemisphere balance, uniformity,
  dead zones, speakers touched). Azimuth is +left / counter-clockwise, matching spatial.ambisonic_encode
  and spatial.export_adm; distance is carried for reporting only. It reads position metadata only —
  no render, no audio — auto-discovers the Object N tracks when no inputs are given, and fails closed
  on an unknown layout, empty keyframes, or more than 128 objects.

The direction encode, the projection-decode speaker basis, the dominant-speaker timeline, and the
dwell / travel / footprint statistics are a new SDK-free src/decode_timeline.h, unit-tested off-REAPER.
Surface is now 186 tools.

[1.4.0] — 2026-07-10

Added

• spatial.orchestrate_sends — automatically wire and reconcile the bed-plus-object send layout
  that feeds an external Dolby Atmos Renderer (or DAPS) input bus. It routes the bed onto renderer
  channels 1…bed-width and each mono object onto one subsequent channel (bed first, matching the roster
  spatial.export_adm and spatial.export_damf emit), taking an explicit object-track list or
  auto-discovering the mono Object N tracks. Reconciliation is idempotent — an existing send already
  on the right channel is reused, a right-source/wrong-channel send is corrected in place, and only
  missing sends are added, so re-running never stacks duplicates. Stray sends with no roster slot are
  reported (and optionally pruned), and it fails closed on the Dolby master limits (at most 118 objects
  / 128 renderer channels). The renderer bus is resized and all sends reconciled inside a single undo
  block; a dryRun preview is the default.
• analysis.send_layout_inspect — read-only inspection and QC of a renderer bus's incoming
  receives: it decodes the send channel assignments, reconstructs the roster, and flags collisions,
  gaps, over-limit object/channel counts, bed-width mismatches, non-mono objects, and unrouted objects.

The send-channel roster and its encode/decode round-trip, the Dolby master constraint validator, the
idempotent reconciliation diff, and the read-back inspector are a new SDK-free src/send_layout.h,
unit-tested off-REAPER. Surface is now 185 tools.

[1.3.0] — 2026-07-10

Added

• spatial.export_damf — author a native Dolby Atmos Master File (DAMF) triad: a .atmos YAML
  manifest (version, presentation, and the bed/object roster with 1-based input IDs), a
  .atmos.metadata YAML sidecar (sample rate plus a flat, ID-keyed event list — bed channels static,
  each object one event per trajectory block), and a .atmos.audio big-endian Core Audio Format (CAF)
  PCM file (interleaved essence, bed channels first then objects in roster order). It reuses the same
  model and stem-render path as spatial.export_adm (so export_adm is unchanged) and enforces the
  Dolby Atmos Master constraints — 48 kHz, a bed no larger than 7.1.2, and at most 118 objects / 128
  channels — failing closed on anything it cannot conform. The room frame is x left→right, y back→front,
  z floor→ceiling.
• analysis.damf_inspect — read-only inspection and QC of a DAMF triad: it round-trips the manifest
  roster, the metadata events, and the CAF header (with endianness warnings), and reports the same for a
  third-party master authored by another tool.

DAMF serialization is a new SDK-free src/damf.h (ADM-spherical → room coordinate map, big-endian CAF
writer, YAML manifest/metadata emitters, and the inspect parser), unit-tested off-REAPER. Surface is now
183 tools.

[1.2.0] — 2026-07-10

Added

• Groove & quantization (MIDI). midi.quantize (snap note starts to a grid with adjustable
  strength and swing; optionally snap note ends too), midi.humanize (seeded, reproducible timing and
  velocity randomization), midi.apply_groove (stamp a per-slot timing/velocity template — from
  midi.extract_groove or a named swing8/swing16 — onto notes), midi.extract_groove (derive a
  reusable groove template from a reference take), midi.transpose, midi.scale_velocity,
  midi.legato, midi.nudge, and midi.stretch.
• Takes & comping. take.add, take.delete, take.crop_to_active, the per-take audio properties
  take.set_vol / take.set_pan / take.set_pitch / take.set_playrate, and items.implode_to_takes
  (stack items across tracks into one multi-take comp item).
• Transport. transport.set_repeat (loop on/off), transport.set_playrate (master play rate), and
  transport.set_metronome (click on/off).

The MIDI groove/quantization math is a new SDK-free src/groove.h, unit-tested off-REAPER. Surface is
now 181 tools.

[1.1.0] — 2026-07-10

Added

• analysis.read_samples — direct, render-free PCM read of a track's item content or a take's
  source via a REAPER audio accessor (no render round-trip). Reports per-channel sample peak, RMS,
  oversampled true-peak, K-weighted level, DC offset, and full-scale clip count, plus an optional
  decimated min/max waveform overview. Bound the window with start/duration; resample with
  sampleRate.
• analysis.accessor_meter — the render-free companion to analysis.meter: per-channel level /
  peak / true-peak / K-level with SMPTE bed labels, L/R correlation, and BS.1770-4 gated loudness,
  computed in-box from accessor samples. A track target measures item content before track
  FX/volume/pan; analysis.meter remains the post-FX render measurement.
• analysis.detect_silence — frame-accurate leading/trailing silence, internal silent gaps, and
  the content-bearing span for trimming, from a windowed cross-channel RMS scan.

New src/audio_accessor.h handles the accessor lifecycle (create → validate → chunked read → destroy)
into the shared metering buffer; the SDK-free helpers (DC offset, clip count, waveform overview,
silence scan) live in src/ambisonic_meter.h. Surface is now 161 tools.

Fixed

• The extension now reports its release version (1.1.0) in the discovery file and build metadata;
  earlier builds reported 0.1.0 regardless of the published release.

[1.0.1] — 2026-07-10

Fixed

• Surround-panner azimuth handedness. spatial.set_source_position and
  spatial.spatialize_stems placed a positive azimuth on the right of a channel bed —
  the opposite of the ambisonic encoder, ADM writer, and metering, where +az = left.
  ReaSurroundPan's in 1 X parameter runs 0 = right … 1 = left, so the panner tools now
  negate the x coordinate when writing to it; a channel bed and an ambisonic scene now agree
  on handedness. (y/z unchanged.)
• analysis.check_deliverable bus measurement. It now measures the requested bus/track
  instead of silently grading the master mix when the target is not the sole route to the
  master (RENDER_SETTINGS source fix).

[1.0.0] — 2026-07-10

Initial public release. Inseglet is a native REAPER extension that exposes a
Model Context Protocol (MCP) interface to the DAW, with immersive/spatial audio and
native ADM (Dolby Atmos) authoring as its focus. 158 tools; live-verified on macOS.

Core control surface

• Project, track, item, take, envelope, FX, send/routing, tempo, and master control
  over a JSON-RPC 2.0 MCP server (protocol 2025-06-18), with bearer-token auth bound
  to a loopback port.
• Granular state setters with consolidated reads; elicitation-gated wholesale state
  replacement with dry-run diffs; an action/command passthrough; and resource
  subscriptions for event-driven workflows.

Immersive & spatial audio

• Channel beds from 5.1 through 7.1, 7.1.4, 9.1.6, and 22.2, with correct LFE handling
  and layout-aware surround positions.
• Ambisonic scene work — encode/decode, order and format conversion, scene rotation
  (yaw/pitch/roll, head-tracking ready), mirroring, beamforming, and distance/room
  placement — via IEM and SPARTA plug-ins.
• Binaural monitoring with live head-tracking.

Native ADM / Dolby Atmos authoring

• Author ITU-R BS.2076 ADM Broadcast-Wave deliverables in-box (bed plus object
  trajectories), with per-object metadata (extent, divergence, importance) and
  round-trip inspection.
• Dolby Atmos Master ADM Profile conformance mode, plus a standalone conformance
  validator for any ADM BWF.

Metering & analysis

• Immersive metering beyond LUFS: loudness, per-channel and correlation metering;
  ambisonic spatial-field analysis (direction of arrival, diffuseness, energy vectors);
  object, stem, dialog, and downmix loudness with timelines; binaural checks; and
  decode-coverage analysis.

Agent ergonomics

• MCP prompts and a deterministic macro-DSL for composing multi-step operations, with
  outcome-level spatial and deliverable-conformance verbs.

Runs on REAPER for macOS; Windows and Linux load-verification are in progress. See
docs/MANUAL.md to get started and docs/REFERENCE.md
for the full tool, resource, and prompt reference.