@photivra/engine 1.5.0 → 2.1.0

package/docs/EXPORT_ACCEPTANCE.md

@@ -1,133 +0,0 @@
-# Simulated Sensor RAW DNG / JPEG acceptance
-
-Release context: **package 1.2.0 candidate / root API 1.2.0**. Subsystem/model versions and dated introduction or measurement records below are independent historical identities; they are not distribution versions. See [developer navigation](DEVELOPERS.md) and [release contract](RELEASE_1_2_0.md).
-
-This record separates implemented export behavior, reproducible independent
-checks and external application acceptance for #16. The files are synthetic
-Photivra captures, never photographs from a physical commercial camera.
-
-## Current implemented boundary
-
-`simulateEnvironmentSensorRawFrame()` executes supplied analytic environment
-queries through the existing optics, wavelength/EQE, local temporal exposure,
-dark/charge/capacity, seeded shot/read noise and ADC pipeline. Its exact native
-post-ADC CFA frame feeds `createPhotographicExportPair()` and the shared #112
-processed path. The serializer does not reconstruct the sensor simulation.
-Attached manual RAW fixtures remain valid low-level serializer inputs, but are
-not the evidence used here for producer-origin acceptance.
-
-Provider transport, visibility, calibration and complete production-stage
-activation are explicitly unverified. Source evaluation and authoritative
-sensor-code production are executed; these are different claims. Bounded production integration is now merged in #178; provider transport and unsupported combinations remain explicit limits, not properties established by a DNG file. Child diagnostics remain accessible in the engine result.
-
-The public export API requires explicit profiles, WB/rendering policy, identity,
-workflow and event time. It copies validated data before asynchronous hashing.
-The commercial app and calling runtime own file downloads. There is no file IO,
-wall-clock sampling, service, new runtime dependency, Adobe SDK or competing
-writer in the engine.
-
-## Reproduce the acceptance files
-
-```sh
-node scripts/generate-producer-export-acceptance.mjs OUTPUT_DIRECTORY
-python scripts/verify-producer-export-acceptance.py OUTPUT_DIRECTORY
-```
-
-The generator uses the existing repository dev compiler to compile only its
-entry point and imported engine/owned fixture modules in a disposable directory.
-It cleans that directory on success or failure. It is not an engine dependency
-or package export. Run from a checkout with its normal dev dependencies installed.
-The Python verifier requires independently installed Pillow, lxml and LibRaw.
-Missing required independent tooling is a failure, not a passing skipped check.
-
-Five 72×48 native pairs are committed in
-`test/fixtures/producer-export-acceptance/`: landscape, both portrait directions,
-inverted landscape and manual WB. Each represents an explicitly identified
-shutter event with distinct artifact IDs; the controlled cases use the same seed
-and physical/noise inputs, so orientation and metadata-WB variations preserve
-all native codes and RAW-data identity. Portrait JPEGs are 48×72, upright JPEG
-orientation is 1, and DNG retains native 72×48 pixels with its orientation tag.
-
-These are explicit #130-derived camera/exposure/seed/units variants. The source
-is an owned three-zone analytic angular environment at 400–500 nm, not the
-canonical Lambertian target or a measured calibration. Optical transmission,
-EQE, uniform aperture response, absent AA/PSF and direct noiseless source
-radiance remain named approximations. Physical sampling uses square 500 µm
-pitch, a declared 0.8×0.6-pitch sensitive aperture, one aperture midpoint and
-one shutter midpoint, with two spectral nodes. Each capture executes 6,912
-provider queries. This is bounded reference evidence, not native camera pixel
-pitch, high-resolution output, measured accuracy or convergence evidence.
-
-The actual generated samples include signed-noise shadows below BlackLevel 64,
-physical clipping and ADC clipping at WhiteLevel 1023. Original RAW integers,
-not an RGB screenshot or later tone curve, supply the DNG. The manual-WB case
-changes AsShotNeutral and JPEG development, while preserving those RAW codes.
-The manifest records exact file SHA-256, RAW-data identity, expected codes/RGB,
-producer model/version/query count, clipping counts and bounded elapsed/heap
-observations. Regeneration can change incidental measurement fields; image
-samples/file hashes remain deterministic for the pinned engine/policy.
-
-## Independent checks
-
-The optional verifier uses Pillow's TIFF directory reader and independent strip
-unpacking for exact RAW codes; Pillow decodes JPEG and standard EXIF. DOM/Expat
-and lxml/libxml2 independently agree on XMP attributes. LibRaw's documented
-opaque-handle C API opens, unpacks and renders every DNG, independently verifies
-native dimensions/white code and reads camera-neutral WB gains. Its emitted
-PPM is decoded by Pillow to verify all oriented output dimensions.
-
-LibRaw API reference: https://www.libraw.org/docs/API-C.html. No LibRaw source,
-private structure layout or version-dependent pointer offset is incorporated.
-Its version is recorded in `independent-validation.json`. Different demosaic,
-tone or display algorithms are not expected to produce pixel-identical images.
-
-Standard capture metadata agree across DNG/JPEG; shared CaptureID/hash remain
-separate from per-resource IDs and native RAW-data identity. Privacy checks
-reject false camera/GPS/serial/host/creator/AI assertions through the existing
-strict metadata contract and verify selected forbidden tags are absent in files.
-Advanced noise hints, calibration matrices/signatures, rights fields and original
-RAW-resource lineage are omitted when no justified mapping exists. See
-[photographic export](PHOTOGRAPHIC_EXPORT.md) and
-[capture metadata](CAPTURE_EXPORT_METADATA.md) for the existing field contract.
-
-## Supported sizes and memory semantics
-
-All existing stage limits remain unchanged: the reference producer,
-reconstruction and full-native export path permit at most **4,096 native sites**.
-The structural attachment ceiling alone does not authorize image execution.
-No 24/45/60 MP export, streaming or resampling is claimed. A tiny output crop
-cannot evade full-native counting. See [RAW frame envelope](RAW_FRAME_ENVELOPE.md)
-and its complete-edge regression tests.
-
-The generator also records 24/45/60 MP unsupported-size preflight requests with
-an empty site payload. They reject before any provider invocation or large image
-allocation. These are measured rejections with deliberately unsupported/missing
-coverage, not typical-resolution memory/performance benchmarks. Before/after
-heap is not peak memory; measured timings are informational. The explicit
-bounded support declaration satisfies #16's supported-limit option without
-claiming high-resolution processing or silently raising safety budgets.
-
-## Closure map
-
-| Requirement | Implementation / evidence | Status |
-| --- | --- | --- |
-| True native post-ADC CFA RAW; exact lossless codes; effective ADC/black/white consistency; no baked processing | `sensor-raw-producer.test.ts`, `raw-output-conformance.test.ts`, producer-derived files and independent verifier | Implemented / independently checked |
-| RAW directly from simulated capture; ISO/gain upstream; same-capture standalone JPEG | `environment-raw-integration.test.ts`, `raw-output-conformance.test.ts`, generated producer pairs | Implemented / bounded origin evidence |
-| Shared capture/provenance/hash; distinct resource/RAW IDs; truthful Photivra EXIF/XMP/IPTC | `capture-export-metadata.test.ts`, `photographic-export.test.ts`, independent file inspection | Implemented / checked |
-| No generative-AI, real-camera, GPS/serial/host/path or fabricated creator/rights claims | Strict input projection and existing negative tests; generated file inspection | Implemented / checked |
-| Native CFA phase, active/default crop, aspect and all orientations | `photographic-export.test.ts`, four generated orientations and LibRaw rendered dimensions | Implemented / checked |
-| As-shot WB metadata does not rewrite RAW; virtual color interpretation is explicit | `photographic-export-correction.test.ts`, same-code manual-WB case and LibRaw neutral-gain interpretation | Implemented / approximate color basis |
-| Deterministic fixture/code/critical metadata acceptance | Committed manifests, hashes and independent TIFF/JPEG/XML/LibRaw verification | Checked |
-| Size/memory boundary and optional runtime/package surface | Existing 4,096-native-site envelope, unsupported-size records, browser/package gates | Explicit supported limits |
-| Adobe RAW interpretation plus independent RAW reader | LibRaw all five pairs; owner Lightroom 9.6 screenshots cover four orientations/as-shot and usable exposure/WB | Reviewed bounded acceptance |
-| JPEG independent decoder/editor/browser acceptance | Pillow all five pairs; owner Photoshop 2026/Chrome landscape screenshots | Reviewed bounded acceptance |
-| Provenance/license review | Existing Apache/DNG notices, owned fixtures and no new dependency; AI-assisted contribution | Owner approved and signed PR #201 |
-| Accurate simulated-RAW documentation | This record and existing metadata/export/frame documentation | Reviewed and merged; reconciled for 1.0 |
-
-## Recorded owner external acceptance — 2026-10-02
-
-[Reviewed PR #201](https://github.com/Photivra/photivra/pull/201) records owner-supplied screenshots: Lightroom 9.6 (build 20260923-1659-a32d844) opens landscape DNG at 72×48, both portraits at 48×72 and inverted landscape at 72×48. Exposure +2.34 brightens the landscape; WB at 6500 K/Tint −78 visibly changes the midtone band. The manual-WB thumbnail uses As Shot and shows a distinct warm band; Adobe's displayed 50,000 K/Tint +28 is its interpretation of synthetic gains, not measured color-temperature accuracy. Photoshop 2026 and Chrome render the landscape JPEG at 72×48.
-
-Independent automated decoding covers all five pairs. User-Mac file hashes and exact Photoshop/Chrome builds were not supplied/reverified; editor/browser screenshots cover the landscape JPEG, not every JPEG variant. These limits are preserved. The committed independent report and `external-review.template.json` describe the earlier automated-only checkpoint; template booleans are not new acceptance evidence. PR #201 merged as `8f5259c4d6b40746173d7842554d416f0138ee27` after owner review/DCO and merge authorization, closing #16. The release audit does not regenerate scientific fixtures, claim additional editor tests or install/buy commercial tooling.
-
-For a future export/model change, rerun the independent reader and relevant Adobe/editor/browser checks against that exact candidate and record hashes/builds. Runtime arithmetic/export schemas remain unchanged in this release preparation.

package/docs/GENERIC_TIER_PRESETS.md

@@ -1,269 +0,0 @@
-# Generic equipment tier reference assets
-
-Release context: **package 1.2.0 candidate / root API 1.2.0**. Subsystem/model versions and dated introduction or measurement records below are independent historical identities; they are not distribution versions. See [developer navigation](DEVELOPERS.md) and [release contract](RELEASE_1_2_0.md).
-
-`resolveGenericEquipmentTierCatalog({ presetVersion: "1.0.0" })` returns immutable
-Consumer, Prosumer and Professional body/lens reference bundles. These are
-Photivra-owned synthetic educational approximations, with unquantified physical
-uncertainty. They are not measured calibration, commercial camera emulation or a
-universal ranking of real equipment.
-
-The tier is a catalog key. Scientific functions consume the ordinary explicit
-profiles; no tier multiplier enters their equations. Sensor format, native raster
-and capture geometry remain caller-owned. All three initial lens forms are 50 mm
-primes with a common f/4–f/16 capability range; prime/zoom and focal range are
-separate asset properties rather than tier meanings. This first catalog contains
-no zoom, macro, OIS or variable focal-length family.
-
-## Exact selection and reference binding
-
-```ts
-import {
-  createGenericEquipmentTierSelection,
-  parseGenericEquipmentTierSelection,
-  resolveGenericEquipmentTierLensProfiles
-} from "@photivra/engine";
-
-const selection = createGenericEquipmentTierSelection({
-  presetVersion: "1.0.0",
-  bodyTier: "consumer",
-  lensTier: "professional"
-});
-const saved = parseGenericEquipmentTierSelection(
-  JSON.parse(JSON.stringify(selection))
-);
-const profiles = resolveGenericEquipmentTierLensProfiles({
-  selection: saved,
-  state: {
-    bodyId: "photivra-consumer-body", bodyVersion: "1.0.0",
-    lensId: "photivra-professional-prime", lensVersion: "1.0.0",
-    focalLengthMm: 50, aperture: 4, focusDistanceM: 5,
-    captureMode: "still", frameRateHz: 0, stabilizationMode: "off",
-    outputWidth: 600, outputHeight: 400
-  }
-});
-```
-
-Omitting `lensTier` pairs matching tiers; all nine pairings are allowed through
-#109's capability resolver. The strict saved-selection parser binds exact IDs,
-preset version and every profile reference, including versions on wrappers for
-subsystems whose native profile lacks a version field. Unknown fields, missing
-references, changed versions and a `latest` alias fail. It returns the canonical
-immutable selection. Output dimensions must be explicit and valid; they do not
-select sensor size. This module does not mutate the POC or production plan.
-
-The optical resolver supports only the explicit still/reference acquisition
-state above. Aperture capability does not authorize extrapolating optical data
-outside f/4. Still stabilization must be off for this optical slice. It exposes:
-
-- radial distortion and lateral-CA channel maps (the CA map already includes the
-  common radial base; do not apply that base twice);
-- normalized illumination vignetting, separate breathing projection scale and
-  a synthetic flat 540–560 nm transmission reference;
-- nine scalar Fraunhofer complex-pupil slices at center, mid-field (10, 6 mm),
-  corner (18, 12 mm), each at signed image-plane defocus −20, 0 and +20 µm;
-- a 550 nm/vacuum clean-lens ghost/veil profile, and separately declared
-  reconstructed-linear geometry/CA/gain correction components.
-
-The 5×5 pupil grids and OPD arrays are independently authored discrete reference
-responses tagged to these exact contexts. They are not an inferred continuous
-lens model or a physically calibrated prediction at arbitrary defocus. No
-interpolation is offered. Normalized PSF shape energy remains one; relative pupil
-throughput stays separate. Pupil clipping and illumination falloff are separate
-contributions with distinct ownership. Their simultaneous production composition
-is not defined here. LoCA, spectral bokeh, onion rings, texture rendering and
-measured MTF are not implemented by these assets.
-
-Correction transforms retain the established destination-to-source semantics.
-Sequential common radial and channel-offset transforms are generic approximate
-compensation; they do not claim the exact inverse of the combined optical CA
-map. Gain is partial and amplifies sampled noise. Correction never recreates
-photons, updates metering, regenerates noise or changes physical capture history.
-The caller must provide any required source prefiltering; a resampler declaration
-is a contract, not proof that prefiltering was performed.
-
-## Curated declarations and tradeoffs
-
-| Explicit body declaration | Consumer | Prosumer | Professional |
-| --- | --- | --- | --- |
-| Maximum standard exposure index | 6400 | 12800 | 25600 |
-| Maximum declared cadence (fps) | 4 | 8 | 12 |
-| Maximum logical sequence frames | 8 | 16 | 24 |
-| Base electronic RMS read noise (electrons) | 6 | 3 | 2 |
-| Data readout duration (seconds) | 0.030 | 0.015 | 0.008 |
-| Pitch/yaw/roll stabilization gain | 0.35 | 0.65 | 0.85 |
-| Continuous AF mode | unavailable | available | available |
-| Highlight-weighted meter/ambience AWB | unavailable | available | available |
-
-ISO selects explicit readout regimes on a finite power-of-two setting grid.
-It is not a photon/noise equation. Full-well/storage physics is not inferred from
-pre-ADC saturation or ADC white. AF uses the same ideal instantaneous actuator
-for every tier; mode availability does not establish faster or more accurate
-tracking. Data readout duration does not establish exposure scan or flash motion
-geometry. Flash capabilities declare front/rear ordinary sync and no HSS; callers
-must resolve actual exposure-window compatibility separately.
-
-The lens declarations intentionally include a tradeoff: Professional has less
-clipping/stray light and lower radial/CA coefficients, but more f/4 falloff and
-peripheral gain cost than Prosumer. Near/far pupil character also differs rather
-than every wavefront coefficient shrinking monotonically. No overall quality or
-bokeh score is produced.
-
-## Acceptance boundary
-
-`test/generic-tier-presets.test.ts` reuses #130's canonical focal length, aperture,
-focus, illumination wavelength and exposure time. It leaves that fixture unchanged.
-A small explicit 9×9 sampled-ramp/noise target tests correction using one capture
-and one noise identity for Off/On. The target is a sampled executor fixture, not a
-radiometrically composed scene or a resolution/MTF claim. Multidimensional pupil
-checks cover all declared field/defocus slices and separate passivity/energy.
-Every advertised ISO setting resolves into an explicit regime. Matched angular
-disturbance, corner radial/CA, off-frame ghost/veil and correction Jacobian/source
-coverage checks consume the existing low-level APIs. The shifted output lattice
-exposes missing support and joint crop rather than extending edges.
-
-`test/fixtures/generic-tier-assets-v1.json` pins the complete canonical manifest
-with SHA-256. Retuning an existing version requires deliberate fixture change and
-review; accepted retuning must introduce a new preset/profile version while
-retaining historical definitions required by saved simulations. This initial
-resolver recognizes only 1.0.0 and rejects unavailable history without fallback.
-
-Initial acceptance was partial; the current #116 disposition is documented in
-[TIER_PRESET_ACCEPTANCE.md](TIER_PRESET_ACCEPTANCE.md). #119 remains the broader
-cross-tier validation ticket. The following inventory describes the initial slice.
-Remaining acceptance includes composed physical scene/corrected-image reports,
-residual CA/distortion and retained-FOV/stretch/edge-response envelopes, controlled
-bokeh/textured backgrounds and spectral effects where modeled, broader body
-stabilization/AF/drive/meter/WB execution cases, and applicable output paths.
-There is no separate tier-specific browser approximation to compare here; the
-public implementation remains browser-safe and uses the ordinary reference APIs.
-No private application acceptance is invented.
-
-Initial tier assets landed at root API 0.105.0; the current root API is 1.0.1.
-This acceptance-only extension does not change runtime/API or asset definitions. Selection schema 0.1.0, asset version 1.0.0, npm package,
-POC and production contracts remain independent. Human scientific/provenance
-review and contribution-specific DCO certification are still required.
-
-## Correction residual and sampled-edge acceptance
-
-`test/generic-tier-correction-acceptance.test.ts` extends #119 with the exact 1.0.0
-assets on a 9×9 synthetic slanted-step target and fixed alternating noise. Correction
-Off/On forks one sampled capture. Independent k1 radial arithmetic, finite-difference
-Jacobians, bilinear weights and polynomial partial-reciprocal gain predict every
-channel sample. These checks quantify sampled interpolation, not optical MTF or a
-radiometrically composed scene.
-
-Independent monotonic scalar bisection recovers ideal coordinates through each
-combined physical radial/CA map. Common green distortion is compensated, while
-sequential partial channel correction leaves a nonzero, reduced CA residual. The
-bounded sampled points require less than 0.1 mm residual displacement; this is a
-regression envelope for synthetic definitions, not measured accuracy or a continuous
-field guarantee. Full 2×2 Jacobians, determinants and principal stretches are checked
-independently. A shifted lattice checks exact missing-source support and the joint
-valid crop. Retained channel pixel-center ray envelopes remain distinct from captured
-pixel-edge FOV; the joint RGB crop cannot expand a channel's ray coverage.
-
-No runtime, root API, schema, preset manifest, package or POC behavior changes.
-This does not complete #116/#119: physical PSF/scene composition, broader body
-execution and applicable export acceptance remain separately tracked.
-
-## Body execution acceptance
-
-`test/generic-tier-body-acceptance.test.ts` executes the exact 1.0.0 body
-assets through existing public APIs using #130's matched optical/exposure state.
-Fifteen cases cover all three tiers:
-
-- Burst times follow the maximum of explicit cadence and exposure duration;
-  a slow shutter cannot overlap ordinary still exposures. Self-timer delay is
-  applied once. Frame seeds are reproducible/distinct within the tested sequence;
-  frame-count limits and unsupported focus bracketing reject.
-- Single AF acquires and holds; focus/release priority gates remain separate.
-  Consumer rejects continuous AF; Prosumer/Professional follow the same known
-  target and require explicit reacquisition after loss. The common ideal actuator
-  does not establish real tracking speed or recognition performance.
-- Uniform/highlight metering uses independently calculated weighted means and
-  target ratios, stays invariant to output crop, and rejects tone-mapped input.
-- Observable pre-WB weighted patches yield independent logarithmic-strength AWB
-  gains. Clipped samples are excluded/reported, unsupported policies reject,
-  and a locked state preserves resolved gains without altering RAW/exposure.
-- Global data-readout declarations retain nonzero duration with zero spatial
-  phase skew under all three shutter mechanisms. These tests do not establish
-  exposure-boundary scans, flash compatibility or rolling-motion behavior.
-
-Assets, runtime contracts and all versions remain unchanged. These are bounded
-synthetic body-policy execution checks, not measured camera calibration or full
-capture/production/output acceptance. Physical scene/PSF composition, applicable
-export wiring and broader stabilization/flash acceptance remain separate work.
-
-## Readout-to-file acceptance
-
-`test/generic-tier-output-acceptance.test.ts` binds the exact body 1.0.0
-readout/signal-chain profiles to an explicitly declared owned 2×2 Bayer/native-still
-test topology. The topology IDs are registered to those required by the assets;
-matching channel names alone is not the binding. Charge expectations, physical
-storage capacity and ideal test-color interpretation remain independent synthetic
-fixture contracts. This tiny raster is structural regression evidence and is not
-an external-editor acceptance file or a realistic pixel sampling model.
-
-Nine cases verify that ISO 100/400 resolve the same base regime and ISO 800 selects
-the explicitly declared high regime. Expected photons/charge and the realized
-stored charge remain unchanged while electronic noise and conversion gain change.
-Independent signed-signal/pedestal/quantization arithmetic predicts each RAW code.
-The producer itself still consumes an explicit regime: changing ISO metadata
-without changing that regime cannot alter its samples. Unsupported ISO settings
-reject in the upstream capability resolver.
-
-Both regimes export through all four orientations. Stored DNG codes match every
-native produced sample; downstream +1 EV preserves codes, capture/noise and RAW
-image identity while changing processed pixels. Repeated JPEG export is deterministic.
-These are genuine producer-to-file checks, with upstream radiometry and producer
-origin verification still explicitly unasserted. They do not establish tier lens
-PSF/scene composition, calibrated color, full production activation or external
-editor/high-resolution interoperability. Runtime, API/schema and asset versions
-remain unchanged.
-
-
-## Supported production and finite pupil scene acceptance
-
-`test/generic-tier-production-acceptance.test.ts` runs every body/lens pairing
-through the authoritative plan 0.7.0 environment route using the exact preset
-transmission values and body readout/ISO profiles. A separately identified fixture
-bridge explicitly binds the selected lens's transmission to its exposure-capability
-profile ID; those two IDs represent different contracts. The test does not retune
-or overwrite the saved 1.0.0 assets.
-
-The owned 2×2 sensor registration, 540–560 nm flat EQE, charge capacity, dark
-current and ideal output color are independent approximations, never facts inferred
-from body tier. All three body selections produce the same expected photo charge
-for a given lens. The lens transmission ratios independently predict photo-count
-ratios; ISO regime changes preserve photo charge. Executed RAW is preserved through
-processing and deterministic paired export. Both consumer manifests retain the same
-immutable executed result. Unsupported wavelength support and stale commitments
-reject before provider callbacks.
-
-Production explicitly omits PSF and declares unity field throughput. It does not
-claim to compose the assets' discrete 550 nm pupil data across a continuous spectrum,
-combine pupil clipping with illumination falloff, or execute every lens effect.
-That omission remains recorded as modeled-zero with evidence and limitations.
-
-A separate finite monochromatic sampled scene evaluates all 27 exact field/defocus
-pupil responses. An independent direct phase sum predicts each normalized PSF tap;
-controlled texture/highlight samples are summed at every tap with metric pitch and
-+Y-up/native +Y-down signs. The same sampled physical response feeds correction
-Off/On, retaining capture/noise identity, full Jacobians and retained ray coverage.
-This coarse sampled scene is neither measured MTF nor a continuous optical model,
-color-fringing prediction or a perceptual bokeh score. No interpolation, sharpen
-pass, hidden quality multiplier or external data is introduced.
-
-See the acceptance disposition for #116 ownership, merged #119 validation and
-scientific limits. Runtime/API/schema/asset/package/POC versions stay unchanged.
-
-
-## Versioned cross-tier reports
-
-The reviewed/merged #119 acceptance layer is [TIER_ACCEPTANCE.md](TIER_ACCEPTANCE.md).
-It links a reproducible three-domain report, normative finite regression envelopes,
-explicit multidimensional pupil/texture and crop/gain tradeoffs, and matched signed
-stabilization/ordinary-flash checks. Existing body, correction and producer-to-file
-acceptance remains part of that evidence. Measured calibration and unsupported
-optical metrics are not invented; runtime/assets and all versions are unchanged.

package/docs/GEOMETRIC_PREPARATION_REUSE.md

@@ -1,141 +0,0 @@
-# Call-local geometry preparation reuse: #43
-
-Release context: **package 1.2.0 candidate / root API 1.2.0**. Subsystem/model versions and dated introduction or measurement records below are independent historical identities; they are not distribution versions. See [developer navigation](DEVELOPERS.md) and [release contract](RELEASE_1_2_0.md).
-
-`calculateGeometricSamplingPlan()` already parses and copies a complete transform
-list at its public boundary. It previously called the public scalar mapping API
-for every destination pixel, reparsing/copying that same list and rebuilding its
-preparation provenance each time. This change uses one private point evaluator
-for the call-owned mapping. The public scalar API still prepares external input
-on every call and uses that same evaluator.
-
-This is a bounded internal optimization against merged main
-`c1374c31473f893310d1ce4f0d5a737aaba43ab5`, not a new prepared/cache API. It consumes
-no pending export/tier/POC draft. Root API 0.103.0, geometry schema/model versions,
-POC, package and production contracts stay unchanged.
-
-## Whole-consumer measurements
-
-[Recorded evidence](benchmarks/2026-10-02-geometric-preparation-reuse.json) contains
-two independent ordinary paired runs and one GC-assisted run. Both source trees
-were built with `tsc`. Context is Node 24.19.0/V8 13.6.233.17-node.51, Linux x64,
-AMD EPYC 9V74 and 9 logical CPUs on shared execution infrastructure. Each run
-records the reference commit and candidate source-file SHA-256; the source/build
-labels are operator-attested, and the fingerprint identifies the runtime change.
-
-The consumers are merged `calculateGeometricSamplingPlan()` and the whole
-`calculateLensCorrectedCapture()` executor: common radial map, channel-specific
-CA, affine map, joint support, actual resampling, and peripheral gain. Both use
-the same owned synthetic capture/noise identity. The sampled linear plane spans
-24×24 mm, with explicit pitch and a separate physical profile normalization
-radius. Nearest sampling is explicit; the bounded compressing coordinate maps
-have no source-prefilter requirement. No prefilter is claimed or fabricated.
-These are real bounded engine consumers, not a copied equation microbenchmark
-or a complete physical production renderer.
-
-| Consumer | Reference ms, ordinary run 1 | Candidate ms, run 1 | Candidate/reference, run 1 | Ratio, run 2 | GC-assisted ratio |
-| --- | ---: | ---: | ---: | ---: | ---: |
-| Plan 1×1 | 0.035 | 0.025 | 0.716 | 0.702 | 0.969 |
-| Correction 1×1 | 0.303 | 0.261 | 0.864 | 0.833 | 0.869 |
-| Plan 5×5 | 0.476 | 0.267 | 0.560 | 0.593 | 0.603 |
-| Correction 5×5 | 3.364 | 2.124 | 0.631 | 0.635 | 0.651 |
-| Plan 17×17 | 5.566 | 2.987 | 0.537 | 0.568 | 0.646 |
-| Correction 17×17 | 39.255 | 24.333 | 0.620 | 0.612 | 0.606 |
-| Plan 33×33 | 22.461 | 11.942 | 0.532 | 0.584 | 0.625 |
-| Correction 33×33 | 155.051 | 100.624 | 0.649 | 0.618 | 0.642 |
-| Plan 65×65 | 90.566 | 59.639 | 0.659 | 0.691 | 0.599 |
-| Correction 65×65 | 640.165 | 468.194 | 0.731 | 0.740 | 0.749 |
-
-There are five measured blocks, alternating reference/candidate order and rotating
-scenario order. The 1×1/5×5 cases use 100 warmup calls and 100 calls per block;
-larger cases use two warmups and one whole call per block. Raw block timing and
-retained-heap deltas are preserved. Final runs occurred after release checks,
-without concurrent task benchmarks/tests. Shared host activity is uncontrolled.
-GC-assisted distributions are separate; do not infer GC causality from ratios.
-
-The two ordinary runs show 26–39% shorter corrected-capture calls at 17×17 through
-65×65. Corresponding sampling plans are about 31–47% shorter. Small cases also
-improve in these runs, but microsecond-scale results have more timing noise.
-These are not portable speed constants, browser frame times, observed product
-traffic, full-resolution throughput or proof of an end-user latency benefit.
-
-The decision is to retain this small internal candidate for review. The result
-justifies removal of duplicate preparation in this concrete consumer, not a
-public mutable prepared engine, extra solver, vignetting/rotation abstraction or
-cross-input cache. #45's POC result is independent and not its evidence basis.
-No browser-motivated public API is introduced; no browser performance claim is
-made. The public implementation remains browser-safe, with the existing package
-and static gates required.
-
-## Preparation, profiling and memory
-
-Separate preparation loops measured about 7–8 µs to parse/copy the three-transform
-mapping; that cost remains at every public boundary. There is no new preparation
-stage or long-lived state. The sampling-plan result still owns its normal mapping
-and all per-pixel diagnostics. Call-local reuse avoids repeating the same work
-for every point. Small grids above include one-off/amortization evidence rather
-than assuming a prepared path is always free.
-
-Separate CPU and allocation sampling used five whole 65×65 corrected captures
-per path. Reference CPU stacks included `prepareGeometricMapping` in 882 of 2,657
-samples inside the consumer (about 33%). Candidate had 1 of 1,779. This is sampled
-inclusive attribution, not exact elapsed time or a claim that preparation costs
-33% everywhere. Recursive finite checks, radial work and other categories overlap;
-do not sum the category counts. Inlining and sampling affect attribution.
-
-V8 Inspector heap sampling, at 32 KiB intervals and requesting inclusion of
-objects collected by major/minor GC, estimated 6,946,184,128 sampled bytes for
-reference and 4,840,744,824 for candidate over five captures (about 30% lower).
-Preparation-inclusive estimates fell from 2,075,859,720 to 656,240 bytes. These are
-statistical allocation estimates, not exact allocation counters, retained live
-memory or GC-pause measurements. Instrumented runs are separate from all timing
-comparisons. No raw inspector profiles/private paths are stored. Raw net retained
-heap deltas are also retained; they are not allocation counts or context sizes.
-
-## Correctness and copy boundary
-
-External prepared objects remain untrusted: every public entry parses/copies the
-current input before evaluating it. There is no persistent context or cache to
-invalidate. The private evaluator is reached only with that call-owned mapping.
-It retains per-point finite checks, every intermediate radial operating-envelope
-check, arithmetic order, full Jacobians/principal stretches and the ordinary
-point-result validation/provenance. Scalar and batch points use the same equation
-implementation. No numerical tolerance is relaxed and no inverse iteration bound
-is changed.
-
-Affine derivative arrays are copied per result point. They must not alias another
-pixel, the returned mapping, the caller or another call. This preserves the prior
-copy semantics despite sharing the internal parsed mapping. Existing masks,
-joint crops, source coverage, prefilter requirements, resampler identities and
-retained-ray/FOV diagnostics are unchanged.
-
-`test/geometric-preparation-reuse.test.ts` compares every point in a fixed 32-case
-corpus (seed `0x43c1374c`) against scalar evaluation, covering all four quarter
-turns, off-center lattices, noncommuting affine/radial order and varied coefficients.
-It also tests mutation isolation, revalidation after prior success, folded/singular
-input and intermediate-domain failure. Existing correction/geometry tests retain
-noise/capture identity and blur/resampling guarantees. The paired benchmark checks
-complete result/provenance equality with the unchanged reference for every measured
-scenario before timing. The canonical #130 fixture is unchanged and is not used
-as performance evidence.
-
-## Reproduction and review
-
-Build an independent reference checkout of the recorded main commit with
-`npm ci` / `npm run build`, and build this candidate checkout. Run sequentially
-without concurrent task checks or benchmarks; replace the reference-root below.
-
-```sh
-node scripts/benchmark-geometric-consumers.mjs --reference-root=/path/to/reference --reference-commit=c1374c31473f893310d1ce4f0d5a737aaba43ab5
-node --expose-gc scripts/benchmark-geometric-consumers.mjs --reference-root=/path/to/reference --reference-commit=c1374c31473f893310d1ce4f0d5a737aaba43ab5 --profile=sample
-```
-
-Repeat ordinary runs separately. Profiling is optional and separate from timing.
-There is no timing CI threshold, new runtime dependency, third-party data, paid
-service or production-stage activation. Human scientific/provenance review and
-contribution-specific DCO certification remain required before merge; measurements
-are not scientific or legal certification.
-
-## Final V1 remeasurement
-
-[Final V1 performance disposition](V1_PERFORMANCE_DISPOSITION.md) records current main #204, fresh paired timing/equivalence, preparation/memory profiles and dense executed environment-to-RAW evidence. PR #205 merged the final reviewed/signed no-further-change disposition; PR #206 merged final conformance. The measured engine commit remains the historical benchmark identity. Earlier candidate and pending-remeasurement statements above are historical.