v1.2.0 — Charts v2, print production, and the conformance channel

[Nizoka]

v1.2.0 — Charts v2, print production, and the conformance channel

Branch: release/v1.2.0 → main
Type: Minor release (additive; one install-time floor raised)
pdfnative: peer floor ^1.6.0 → ^1.7.0
Node: >=22 (unchanged)

Summary

Tracks the pdfnative engine's 1.7.0 release and delivers the one capability
this package had promised in writing (docs/CHARTS.md: Charts v2 "reach this
package as new ChartProps fields"), plus print production, the PDF/A
conformance diagnostics channel, seven new lint rules (18 → 25), and the
quality backlog deferred from the 1.1.0 review (cacheControl/etag on
renderToResponse, ErrorOptions/cause, eslint-plugin-react-hooks,
validateSpec fuzzing, the PDF/UA round-trip test, raised coverage
thresholds). No public API was removed or changed.

Install-time floors

Requirement	1.1.0	1.2.0	Why
pdfnative (peer + devDep)	^1.6.0	^1.7.0	The four new chart kinds, layout.print and the diagnostics channel do not exist before 1.7.0; ChartPropsCoversChartBlock cannot even compile against 1.6.0 typings. doctor() reports a 1.6.x engine with a dedicated upgrade message (second capability probe: validatePrintOptions).

Changes

src/components.tsx

• ChartProps + 5 projections (axis2, xAxis, dataLabels, labelStride,
  labelRotation), each typed as ChartBlock['…'] so future inner-shape
  changes absorb automatically. The ChartType widening (5 → 9), axis.scale
  and ChartSeries.xValues/yAxis flowed through the existing projections
  with zero code. The ChartPropsCoversChartBlock compile lock forced all of
  this on peer bump — exactly its design intent; its docblock now records the
  first firing.
• DocumentProps.print?: PrintOptions — new sugar prop (watermark/attachments
  precedent).
• TableProps.cellVAlign docblock corrected: omitted ≠ 'top' (the engine
  keeps its historic baseline placement; behaviour was always right, the
  JSDoc was not).

src/reconciler/serialize.ts

• Chart case forwards the five new keys through compact() (no compile-time
  lock covers this list — the extended golden snapshot is the guard).
• resolveLayout folds print into layout.print. Invariants re-pinned by
  tests: explicit layout wins; no sugar + no layout ⇒ layout === undefined.

src/spec/ (types, validate, compile, schema)

• DocSpec.print (+ KNOWN_FIELDS, forced by KnownFieldsAreExhaustive;
  specToElement passes it). ChartSpecBody = ChartProps inherited everything.
• Schema: chart builder covers the full 1.7.0 surface (9-kind enum in engine
  union order, series xValues/yAxis, axis.scale, axis2, xAxis,
  dataLabels oneOf, labelStride ≥ 1, labelRotation 0–90); print on the
  doc-spec schema (boxes as 4-number arrays, userUnit 1–75000, bleed
  exclusiveMinimum 0); metadata description names trapped; render-options
  layout description enumerates print/outputIntent/strict/onDiagnostic
  and the four new viewerPreferences keys, with the explicit note that
  onDiagnostic is function-valued and JSON-unrepresentable (strict: true
  is the JSON-safe switch). Link/svg opts descriptions completed
  (fontSize/color, alt) after the exhaustiveness audit.
• Deliberate non-additions: outputIntent, strict, onDiagnostic are NOT
  DocSpec top-level fields — non-JSON or niche; they remain reachable through
  the layout passthrough. Every added top-level field is permanent API.

src/registry.ts + src/lint.ts (18 → 25 rules)

• L_CHART_LOG_SCALE, L_CHART_X_AXIS, L_CHART_LABELS (errors) — every
  Charts v2 constraint verified against the engine's validate() in
  pdf-chart.ts, reported pre-render. L_CHART_CATEGORIES now skips
  positional-axis charts (mirrors engine 1.7.0 exactly — the old behaviour
  would have been a false positive).
• L_PRINT_BOXES (error) — delegates to the engine's own
  validatePrintOptions (with the engine's real PG_W/PG_H defaults) in a
  try/catch; the finding carries the engine's message verbatim. Zero
  duplicated geometry rules, zero drift, by construction.
• L_VIEWER_PRINT_RANGE (error), L_OUTPUT_INTENT_IGNORED (warning — the
  engine silently ignores outputIntent without tagged),
  L_TAGGED_FORM_FONTS (warning — mirrors PDFA_UNEMBEDDED_FORM_FONT).
• Rejected rule candidates, with reasons: "scatter without xAxis" (does not
  throw — scatter defaults to a linear axis; verified in engine source),
  "axis2 without a right-bound series" (silently ignored, harmless), "CMYK
  image under PDF/A" (would require parsing JPEG bytes in a pure linter; the
  engine's PDFA_DEVICE_CMYK_IMAGE diagnostic covers it at render time).

src/render.ts + src/response.ts

• Streaming pre-flight (exhaustiveness-audit finding): the engine rejects
  <TableOfContents> and {pages} templates on its streaming path, but
  checked inside the generator — renderToResponse (streaming by default)
  failed mid-response, after status and headers were sent. renderToStream
  now calls the engine's validateDocumentStreamable eagerly; all streaming
  entries throw catchably at call time. Two regression tests pin it.
• renderToResponse: opt-in cacheControl and etag (string verbatim, or
  true to derive a strong FNV-1a validator from the bytes — implies
  buffering). Defaults byte-identical when unset.

src/doctor.ts, src/manifest.ts, src/core-bridge/index.ts, src/index.ts

• REQUIRED_ENGINE = '1.7.0'; graded probe (validatePrintOptions → ok;
  estimateChartHeight only → "1.6.x — this release needs >= 1.7.0"; neither
  → missing/older).
• contract.engine: '^1.7.0'; new manifest entrypoint setDeflateImpl.
• Bridge: + validatePrintOptions (probe + lint delegate), PG_W/PG_H,
  validateDocumentStreamable, setDeflateImpl; − buildDocumentPDFStream
  (imported since 1.1.0, used nowhere — dead surface in the "small, auditable
  slice"). New type re-exports: PrintOptions, PrinterMarksOptions,
  PageBox, CustomOutputIntent, PdfDiagnostic, PdfDiagnosticCode,
  PdfDiagnosticHandler, PdfColors.
• Barrel: + setDeflateImpl (closes the browser-compression asymmetry with
  initNodeCompression) and the eight new types.

src/errors.ts, src/hooks.ts, eslint.config.js

• PdfReactError/PdfStructureError accept ES2022 ErrorOptions
  (error.cause reachable; JSON envelope unchanged — asserted key-exact by a
  test).
• eslint-plugin-react-hooks@7 (recommended-latest, flat config). Its two
  new rules flag the deliberate latest-value-ref and loading-flag patterns in
  hooks.ts; each of the four sites carries a per-site
  eslint-disable-next-line with the reason, keeping the rules armed for
  future code instead of a rule-wide opt-out.

scripts/generate-pdfa-corpus.mjs + scripts/validate-pdfa.mjs + .github/workflows/verapdf.yml (new)

• The veraPDF conformance gate, ported from the ecosystem's most mature
  implementation (pdfnative-cli): an 11-file manifest-driven corpus rendered
  through the built package (9 positives across pdfa1b/2b/2u/3b through
  both authoring doors, incl. Charts v2 dual axis, print production, and a
  form with embedded fonts; 2 negative canaries — plain text and the AcroForm
  content path, both without embedded fonts, ISO 19005-2 §6.2.11.4.1).
  Runner: strict single-<validationReport> parsing,
  6-outcome taxonomy, fail-closed VERAPDF_REQUIRED=1, .bat-through-shell
  quoting (CVE-2024-27980), raw-report artifacts. CI blocking (user
  decision; engine/CLI precedent), pinned veraPDF greenfield 1.30.2 with
  installer SHA-256 verified before java -jar; the same gate runs
  pre-publish in publish.yml before the SBOM.
• samples/layout/page-setup.tsx fixed: it claimed PDF/A-2b without fonts —
  non-conformant, and invisible to L_TAGGED_NO_FONTS (claim via
  RenderOptions.layout). Now embeds Noto Sans and documents the blind spot.
• .gitignore: root-anchored /*.pdf for stray sample outputs.

Tier 5 — visual verification for vision agents (new)

• docs/AGENT_CONTRACT.md gains the post-render tier (extractText →
  validatePdfUA/veraPDF → rasterize + look); docs/RECIPES.md gains the
  matching recipe (also repairing docs/LINTING.md's dangling veraPDF
  cross-reference); llms.txt documents it for agents;
  samples/agent/visual-verify.tsx runs it end to end with graceful
  degradation to the tier-4 geometry report when no rasterizer
  (pdftoppm/mutool) is installed. No dependency added — the rasterizer is
  deliberately external.

npm discovery metadata (package.json)

• description rewritten for the 1.2.0 surface (charts, print production,
  veraPDF-validated PDF/A, the agent surface); keywords 42 → 68, cluster-
  ordered — the ecosystem's shared print/charts/PDF-A vocabulary plus
  agent-discovery terms (llm, agent-tools, and the react-original
  docspec/capability-manifest/verapdf), and the documented
  react-pdf-alternative anchor. Deliberately excluded: signature/PAdES
  terms (authoring-only boundary), MCP client names (a library is not a
  server), factur-x (no dedicated recipe). Nothing pins these fields
  (verified); every keyword maps to a documented capability.

Version plumbing

src/version.ts → 1.2.0; package.json; CITATION.cff (+ date-released: 2026-08-26). Governance spec (.github/ai-governance.json / governance.ts)
stays at 1.1.0 deliberately: the policy content is unchanged, and bumping
a spec version without changing the spec is drift in the other direction.

Samples & tests

• New samples (each run end-to-end, valid PDF verified — four in total):
  samples/charts/charts-v2.tsx (8 charts: stacked, area, scatter, time axis,
  dual axis, log scale, data labels, label rotation),
  samples/layout/print-production.tsx (bleed + marks, trapped, duplex /
  printPageRange / numCopies), samples/quality/diagnostics.tsx
  (onDiagnostic collector, strict throw, and the lint tier side by side),
  and samples/agent/visual-verify.tsx (the tier-5 visual loop, described
  below). samples/README.md rows added.
• Tests: 226 → 292 (18 files). Charts v2 round-trips (serialization,
  DocSpec parity, real renders for all four new kinds + time/log/dual-axis),
  print sugar folding + layout === undefined re-pinned, one assertion per
  new lint rule (with clean-case counterparts and the positional-axis
  regression), cache-validator behaviour (deterministic ETag, merge-last),
  cause in-process but out of the envelope, streaming pre-flight, extended
  golden snapshot (+81/−6, the diff read before -u), new
  tests/fuzz-validate.test.ts (seeded PRNG, 700 malformed/mutated inputs,
  never-throws + well-formed reports), new tests/pdfua.test.tsx (render
  tagged → engine validatePdfUA → no violations; the 1.1.0-deferred
  round-trip).
• Coverage thresholds raised 85/80/85/85 → 90/84/92/90 (measured:
  95.0/89.8/97.8/95.9).

Docs & governance

• src/fonts.ts — correctness fix found by the panel (see round 2):
  resolveFonts emitted fontRef without the leading slash; the engine
  writes it verbatim into content streams as a PDF name, so every document
  produced through the documented font path (resolveFonts /
  options.fonts) was malformed (BT latin 10 Tf). Now normalized
  (/-prefix), with the sample and test expectations updated and the PDF/UA
  round-trip upgraded to the embedded-fonts configuration it originally
  could not pass.
• Updated: README.md (component table, page-furniture section, "Upgrading
  to 1.2"), llms.txt (full 1.2.0 surface, 25-rule table, streaming
  restriction, setDeflateImpl), docs/CHARTS.md ("Charts v2 — a promise
  kept"), docs/LINTING.md (25 rules, the diagnostics channel as "tier 4½"),
  docs/SERVER.md ("What cannot stream", caching), docs/KNOWLEDGE_BASE.md
  (probes, test map, design boundaries incl. the newly-written table-builder
  exclusion rationale), docs/AGENT_CONTRACT.md, docs/RECIPES.md (PAdES
  LTV pointer), ROADMAP.md (1.2.0 shipped; fixed the stale "five" count),
  CLAUDE.md, .github/copilot-instructions.md.
• release-notes/v1.2.0.md + this draft.
• An upstream issue draft blaming the engine's validatePdfUA was prepared
  during test-writing and withdrawn in round 2: the panel's factuality
  validator could not reproduce it, and the arbiter proved by bisection that
  the failure was this package's own fontRef bug (above). There is no
  engine bug to report.

Validation

npm run typecheck:all   ✔ (src + tests + samples)
npm run lint            ✔ (now includes eslint-plugin-react-hooks)
npm test                ✔ 18 files, 292/292
npm run test:coverage   ✔ 95.04 / 89.80 / 97.83 / 95.92 (thresholds 90/84/92/90)
npm run build           ✔ 8 dist artifacts, tree-shake probe ok (version-only
                          bundle 3354 bytes, no reconciler), 'use client'
                          restored on client bundles, absent from root
npm audit --omit=dev --audit-level=high   ✔ 0 vulnerabilities
New samples executed manually             ✔ 4/4 produce valid PDFs (incl. visual-verify
                                            with graceful no-rasterizer degradation)
npm run corpus:pdfa                       ✔ 11 files + manifest (2 negative canaries;
                                            engine diagnostics observed as designed)
First blocking CI veraPDF run             ✔ after round 4: 9 PASS + 2 XFAIL expected
                                            (round-4 XPASS event resolved below)
node scripts/validate-pdfa.mjs            ✔ canaries green; SKIPPED locally (no veraPDF/Java
                                            on the dev machine — exit 0 is a skip, not a pass;
                                            CI runs the same corpus blocking + fail-closed)

Adversarial review

Round 1 — two independent exhaustiveness audits (blind to each other)

Audit A walked the engine's 445-export barrel down; audit B walked every
parity chain (JSX ↔ DocSpec ↔ schema ↔ manifest) field-by-field up. B's
verdict: parity complete, two cosmetic schema-description gaps. A's verdict:
"not exhaustive — narrowly", eight findings. All fixed or dispositioned:

Finding (auditor)	Resolution
Streaming rejects <TableOfContents>/{pages} inside the generator — breaks renderToResponse mid-response, undocumented (A)	Fixed: eager validateDocumentStreamable in renderToStream; documented in SERVER.md, llms.txt, JSDoc (the "feature-equivalent" claim corrected); 2 regression tests
setDeflateImpl unreachable — browser layout.compress silently produced larger output (A)	Fixed: bridged + barrel + manifest + llms.txt
Table-centric builder (buildPDF/PdfParams) excluded without recorded rationale (A)	Fixed: rationale written into KNOWLEDGE_BASE §8 ("One builder, not two")
PdfColors not re-exported while layout.colors is documented (A)	Fixed: type-only re-export
cellVAlign JSDoc claimed default 'top'; engine keeps historic baseline (A)	Fixed: JSDoc corrected (behaviour was always right)
RenderOptions.fonts JSDoc omitted renderToResponse from the honored list (A)	Fixed
svg schema description omitted alt; link opts omitted fontSize/color (A + B, independently)	Fixed: descriptions + link properties completed
Dead bridge import buildDocumentPDFStream (A)	Fixed: removed

Rejected findings, with reasons:

Rejected	Why
Surface buildDocumentPDF (string twin), buildDocumentPDFStreamPageByPage, wrapText, PAGE_SIZES, font-registry introspection helpers (A: "not-findings")	String output is a legacy shape; the page-by-page stream is superseded by the true stream; page sizes are reachable numerically; registry helpers are dev/test utilities. Concur with the auditor's own disposition.
Schema over-constrains ticks to integer >= 2 (B, cosmetic)	Deliberate: the engine treats fewer than 2 ticks as meaningless; a tighter authoring schema that only excludes nonsense inputs is a feature, and validateSpec does not enforce it (opts objects stay open).

Round 2 — post-implementation panel: conformity/standards, factuality, and an arbiter

Two independent validators (one on conformity and OSS/industry standards, one
on factuality — every claim executed, not read), then a third agent arbitrated
the legitimacy of every finding, resolving the one head-on contradiction by
execution and bisection.

The headline: the factuality validator could not reproduce the upstream
validatePdfUA bug the test-writing pass had reported (and drafted an issue
for), while the original observation demonstrably happened. The arbiter ran
both sides' reproductions and proved a third explanation: the engine has no
bug — this package's resolveFonts emitted fontRef without the leading
slash, so every document rendered through the documented font path was
genuinely malformed (BT latin 10 Tf — a keyword where ISO 32000 requires a
name). The validator was the messenger. One bisection table settled it:
fontRef: '/F3' valid, 'latin' invalid, '/latin' valid, 'F3' invalid.

Finding (validator)	Ruling	Resolution
Issue draft's central claim does not reproduce (factuality) + pdfua test docblock repeats it	Legitimate symptom, wrong diagnosis (arbiter, by execution)	Fixed: resolveFonts normalizes fontRef; issue draft withdrawn (no engine bug); pdfua test upgraded to the embedded-fonts + pdfa2b configuration and its docblock rewritten; sample + test expectations updated; CHANGELOG entry with re-render guidance
Release work uncommitted at review start (conformity)	Legitimate	Fixed: committed to release/v1.2.0 before round 2 closed (this table lands as its own commit)
package-lock.json root version still 1.1.0 (conformity)	Legitimate	Fixed: lockfile regenerated (both version fields 1.2.0)
Coverage/bundle figures drifted by a hair (both validators)	Legitimate	Fixed: re-measured after the round-2 fixes and pasted verbatim (95.04/89.80/97.83/95.92; 3354 bytes)
AGENTS.md rule 1 contradicts the sanctioned test-side engine import (conformity)	Legitimate	Fixed: rule 1 now scopes the bridge requirement to src/ and names the tests exception
ROADMAP presented a paraphrase as a quotation (factuality)	Legitimate	Fixed: quotation restored verbatim
CHANGELOG headings carry no ISO date (conformity)	Rejected (arbiter)	House style since 0.1.0 is ## [x.y.z] — tagline; dating only 1.2.0 would be inconsistent and break the GitHub anchors the release notes link to. If dates are ever wanted: all headings + both anchors in one housekeeping change, not mid-release.
strongEtag is FNV-1a, not cryptographic (conformity)	Rejected — no action (validator's own disposition, arbiter concurs)	Already documented as change-detection; HTTP strong validators need byte-identity semantics, not a cryptographic hash.

Round 3 — full-release control panel (post-veraPDF/vision work)

Two fresh validators re-controlled the entire release (all three commits):
validator A on conformity, project philosophy and breaking changes; validator
B on factuality (every claim executed, both ports diffed against their
pdfnative-cli source) and 2026 OSS standards. A third agent arbitrated.

Verdicts: A — zero breaking changes (the two judged cases: the
resolveFonts byte change corrects spec-violating output and is disclosed
with a re-render note; new lint rules add at most a warning on pre-existing
documents, report.ok preserved) and philosophy conformant on all 8 golden
rules. B — factual: both ports are logic-identical to the CLI originals
(prose-only deltas), the SHA-256 pin matches byte-for-byte, fail-closed
behaviour was executed and confirmed (exit 3 under VERAPDF_REQUIRED=1
without veraPDF), and the corpus/canaries behave exactly as documented.

Finding (validator)	Ruling (arbiter)	Resolution
CONTRIBUTING's local Linux recipe executed the installer without the SHA-256 check CI performs (A)	Legitimate	Fixed: sha256sum -c line added — local and CI now verify the same pinned artefact
PR draft mis-rounded branch coverage to 90.0 in two lines while its own validation block said 89.80 (B)	Legitimate	Fixed: 89.8 everywhere
"4/4 samples" vs three named in the Samples bullet (B)	Legitimate	Fixed: visual-verify.tsx named in the bullet
KB's "one place a test may import pdfnative directly" overstated — font-data subpaths are a documented consumer pattern (A)	Legitimate	Fixed: reworded to "the engine's API directly"
failedRules regex in the validator is attribute-order-dependent (A)	Legitimate — keep verbatim	The verdict parse is independent and an emptied listing prints an explicit marker; fixing only here would diverge the ecosystem port. Noted as an upstream follow-up (fix in pdfnative-cli first, re-port everywhere).
2> >(tee …) flush race in the workflow (A)	Rejected	Exit code travels via PIPESTATUS; per-file stderr is independently written and uploaded; identical construct runs in production upstream CI.
L_PRINT_BOXES could surface a raw TypeError on a 1.6.x peer (A)	Rejected — premise wrong	The call already sits in the rule's try/catch, and a 1.6.x ESM peer fails at module load before lint runs; out-of-contract regardless (doctor() reports it).

Round 4 — first blocking CI veraPDF run: the XPASS guard fired, and was right

The very first CI run of the new gate failed with
XPASS form-pdfa2b.pdf (negative canary ACCEPTED) — exactly the fatal guard
working as designed, against us: the manifest's expectation was wrong,
not the validator. Analysis: the engine's PDFA_UNEMBEDDED_FORM_FONT
diagnostic is conservative — it fires because the AcroForm /DA references
/Helv — but in this corpus entry (embedded fontEntries, no field values)
no text actually renders through /Helv, so veraPDF 1.30.2 correctly accepts
the file. ISO 19005 requires embedding for fonts used for rendering.

Resolution: form-pdfa2b.pdf became a positive (empirically CI-proven,
and now valuable coverage: forms + embedded fonts pass under PDF/A), and a
genuine negative replaced it — form-nofonts-pdfa2b.pdf (form + text with no
embedded fonts, the AcroForm content path of §6.2.11.4.1, same
proven-rejection class as nofonts-pdfa2b). Corpus 10 → 11 files, still 2
negative canaries; all counts swept.

Upstream observation for pdfnative-mcp: its corpus carries the same
form-pdfa2b / expectCompliant: false expectation ("veraPDF reports
6.2.11.4.1") that this CI run disproved for the no-values case — and its
veraPDF workflow is advisory (continue-on-error), so an XPASS there would
currently pass unnoticed. Worth re-checking when mcp's gate goes blocking;
note their entry fills field values through /DA /Helv, which may genuinely
render through the unembedded font — the verdict may legitimately differ.

Backward compatibility

Change	Impact
Peer floor ^1.6.0 → ^1.7.0	Install-time only; npm refuses a stale engine at install, doctor() explains at runtime. Not an API change.
renderToStream validates streamability eagerly	Previously the same documents failed at first pull (or mid-response); they now fail earlier, catchably, with the same engine message. No working document is affected.
etag: true implies buffering	New opt-in option; no existing call sites.
Engine 1.7.0 rendering changes (RTL, /ToUnicode, x-label stride)	Inherited byte-level differences documented under Compatibility in the CHANGELOG; wrong-before-correct-now per the engine's release notes.
Coverage thresholds raised	CI-internal.
Everything else	Additive. Documents that use no new feature compile byte-identically (layout === undefined invariant re-pinned; golden snapshot diff reviewed).

Out of scope (by design)

• Engine 1.7.0's PAdES LTV/signing/timestamps/DSS, PdfModifier.updateMetadata,
  form fill/flatten, merge/split, extraction — byte-level post-processing
  (golden rule 7); docs/RECIPES.md points at the engine, now including the
  LTV ladder.
• The MCP server, React Native renderer, incremental compilation — ROADMAP
  "Later" (user decision on this release's scope).
• Upstream repo updates that belong to Nizoka/pdfnative:
  docs/assets/ecosystem.json (packages.pdfnative-react.version/pin) and
  docs/data/surfaces.json React cells must move to 1.2.0 / ^1.7.0 as part
  of the engine repo's release train, or its verify:docs will fail.
  Optionally, an ergonomics suggestion for the engine: validate or normalize
  fontRef format at the API boundary, so a missing slash fails loudly
  instead of corrupting output.
• Ecosystem-wide follow-up: the veraPDF runners' failedRules regex (here,
  pdfnative-cli and pdfnative-mcp — all verbatim ports of the same
  original) assumes veraPDF's <rule> attribute order; harden it upstream in
  pdfnative-cli first, then re-port, so the three repos stay identical.
• ~21 open dependabot branches — separate housekeeping.

Self-review checklist

• 1. Runtime pdfnative imports only via src/core-bridge/index.ts; src/types.ts type-only; peer dependency intact (test-pinned).
• 2. No CSS layout model; no new composite (<Section> remains the only tag: null entry); print is a prop, not a component, because it is page furniture.
• 3. Reconciler contract untouched (React 19 / react-reconciler@^0.31 / @types@^0.32).
• 4. Strict TypeScript, no any; lint green including the new hooks plugin.
• 5. 'use client' only via the client subpath; src/spec/ stays pure; postbuild asserts both directions.
• 6. Full parity JSX ↔ DocSpec ↔ registry ↔ schema for every new capability; src/version.ts bumped; version triple + $ids test-pinned.
• 7. Authoring only — every 1.7.0 post-processing capability excluded and the exclusions re-audited one by one (round 1).
• 8. Draftsman, not submitter: this PR draft and the issue draft await human submission; verify:issue passes; no autonomous GitHub writes occurred.

Compliance report

Field	Value
no_new_runtime_dependency_confirmed	Yes — dependencies is still exactly ["react-reconciler"] (test-pinned). eslint-plugin-react-hooks is a devDependency.
reproduction_command	npm run typecheck:all && npm run lint && npm test && npm run build
reproduction_result	All green: 292/292 tests, coverage 95.0/89.8/97.8/95.9, 8 dist artifacts verified.
duplicate_search_performed	Yes — CHANGELOG, ROADMAP, release notes and open drafts checked; no existing 1.2.0 work. A drafted upstream issue was withdrawn after the panel disproved its central claim (round 2).
affected_packages	pdfnative-react (this repo). Follow-ups noted for pdfnative (ecosystem manifest, surfaces.json — doc alignment only; no engine bug).
identity_reminder_shown	This draft must be reviewed and submitted by a human under their own GitHub identity. No agent will open the PR; no issue remains to submit.