[twoimo]

Alden now searches complete retained UTF-8 TXT/Markdown originals, including
matches beyond the parser's16,000-byte projection and across that boundary.
A separate reversible original index retains general text/vector policies.
Current project/target/version/raw/config/capture bindings admit candidates
before limits, and selected query-centered excerpts are rebuilt from verified
original bytes. Their proof separates whole-input/projection hashes, delivered
text and exact or disclosed covering original byte ranges. Initial parser
coverage remains independent. Capture commits before optional derived indexing;
index failures report a separate failed stage, and lookup never builds an index.

Original ranks replace parser ranks for the same file, with one interleaved
lexical budget. Selected-file term contributions center multi-term excerpts so a
common head match cannot hide the discriminating tail. Cached scores remain
advisory. Final source validation inherits cancellation and checks again before
returning. Typed knowledge authority rebuilds the selected original excerpt and
raw range without rerunning search. Voice keeps its actual match within the
existing fact limit. Unicode query terms preserve accented letters and combining
marks while retaining the prior ASCII/Hangul particle rules.
Korean compound-particle queries append roots for six closed suffixes only
when the existing24-term budget has spare slots. All legacy terms remain the
same prefix; source identity, scope and the graph literal filter stay intact.

Explicit project lookup prompts separate a literal project/source wrapper and
requested output format from the content search terms. Existing host intent and
project resolution run first; unknown syntax keeps the original query. A bounded
hash-only trace binds request and retrieval text without replacing user messages
or widening permissions.

Exact retained-original body windows, desktop first/next/reset and the read-only
alden_knowledge_source_body MCP tool remain available with their prior bounded
payloads, no original-path fallback and current source authority checks.

Original-file reads also preserve keyboard focus during asynchronous loading.
The busy guard rejects repeat activation while accessible disabled states keep
the focus target. At EOF, Next transfers focus to Reset only when it is still
focused; another control selected during the read keeps focus. Source and byte
verification, stale-result fencing and callback contracts remain intact.

alden-cli becomes the canonical internal command across execution, packaging,
help and completions. The existing compatibility executable and persisted app,
data and permission identities remain supported. The interface uses Hallmark's
warm navy direction, with explicit partial, stale and unknown states.

Knowledge requests bind exact model routes, source authority and retained
versions. Typed requests permit at most two model hops and one authorized source
read, with bounded nonstream responses and one final commit. Malformed tool
calls, truncation, source changes and incomplete replies stay typed failures.
Rejected typed local completions retain bounded HTTP finish/shape/call-count
observations before strict parsing. Duplicate calls still cause zero source
reads. Sanitization excludes content, arguments and arbitrary wire values;
diagnostic compaction respects the existing4 KiB receipt limit after query/
attachment context and late failure-code fields are added.

Native inventory and observed local endpoints are validated separately from
router aliases; partial router availability does not create remote health or an
automatic fallback.

The new explicit file-set adapter reads declared original documents rather than
treating Spark's derived index as original-file acquisition. It retains stable
file IDs, exact original bytes, revisions and structural locations; equal-byte
copies share a blob without merging identity. Acquisition failures cannot prune
the prior set. Verified-root descriptor reads and pre-publication source/target
guards reject races. PPTX linked notes are read with real slide/part/paragraph
positions. Coverage is explicit: Office/PDF/conversion extraction stays
conservatively incomplete, even when original-byte capture is complete.
Manifest, limits and evidence.

Graph and terminal changes retain source provenance through GBrain/open-ontologies
integration and normalization. The ordinary relationship budget stays 512. A
bounded integer-coordinate neighbor table improves the pure pin solver while
preserving results; other coordinates retain the original lookup. Worker faults
release ownership and invalidate late callbacks. A subsequent explicit solve
creates a fresh worker. The terminal f(x) counts deduplicated permitted stored
document transitions by UTC interval and binds its full snapshot checkpoint
separately from history paging. It does not infer original-source completeness
or live intake from node modification dates.

Validation:

• Rejected knowledge replies at b848633: 39 new regression methods /
  254 subcases pass with generated sources and mocked transport; the normal
  repository unittest module command also passes39 methods. The copied prior
  source detects absent observations. 106 related existing methods were
  accepted with only owned parser/image/CLI children and one ephemeral HTTP
  fixture admitted. Tests cover privacy, zero-read duplicate rejection,
  persisted failure/replay, positive tool/final/no-call cases and4 KiB pressure
  from query provenance, attachments and late errors. Optional diagnostic shapes
  or older observations can be explicitly omitted without changing primary
  evidence, actual calls or protocol acceptance. Exact-source CI
  passed 4/4 jobs. These are plumbing tests, with no real model or original-data tool execution or local
  app startup. The prior local7b4a57f candidate and installed0.3.219 are preserved;
  this change has no new local installed-runtime proof.
  Diagnostic contract.
• Actual offline fixed27B diagnostic at 7b4a57f: the existing pinned
  checkpoint14 files/18,219,586,850 bytes match before and after;34 runtime
  distribution versions and5,751 RECORD hashes were verified before imports.
  Two real native loads/generations ran under the canonical exclusive lease
  and memory/state guards. Strict single-tool decision acceptance:0/2.
  Attempt001's specific rejection cause is unknown because failure text was
  not retained. Only diagnostic recording changed before a separately reviewed
  second attempt. Attempt002 stopped normally with382 prompt/34 generation
  tokens but emitted two identical empty tool calls. Exact token replay matches
  SDK text; the single-call guard correctly rejects the whole response without
  dedup/drop/repair. Both runs executed zero source reads/SQL/facts/final answers.
  Genuine source previews separately fail CANTOPEN: current WAL databases have
  no sidecars under the physical no-outside-write diagnostic profile. Owned
  generated controls support that condition; full source/final remains not
  admitted. No source write/checkpoint/immutable/nolock/raw-copy/backend bypass
  occurred. The2-attempt budget is exhausted. Source/helper/protected state,
  installed85-file app and Alden OFF are preserved. Independent native public
  receipt review agrees with NOT_ACCEPTED. This is native decoder-component
  failure evidence, not successful product HTTP/voice/installed integration.
• Korean compound queries at 7b4a57f: 323/323 local methods pass,
  including28 new methods/173 subtests and11 existing Unicode methods;
  1,400 generated combinations retain the legacy prefix/cap/NFC behavior.
  The parent deny-all harness initially blocked one existing owned socket
  fixture; that test passed with only its ephemeral listener admitted, without
  changing product or existing test code. Exact-source
  CI40794626319
  passed 4/4 jobs, including unsigned arm64 package/runtime checks.
  After CI, a fresh40-file copy of the same22 retained originals was admitted
  against358 source pins and32 exact readiness files. A source-only annotator
  fixed12 questions/31 quotes from6 previously excluded originals before
  implementation; baseline scores stayed sealed until source freeze.
  Paired retrieval at2or4 entities/400-character original excerpts/8,000 context
  characters remained unchanged: source inclusion@1 10/12, @5 11/12,
  delivered source 11/12, all annotated quote ranges 2/12. No metric or
  individual case lost acceptance, but no measured gain on this reference
  packet is claimed. Both40-file states remain byte/mode/path identical;
  source/version/hash checks pass, outside-scope facts and guarded network/
  embedding calls are zero. These author-selected references may have a phase
  directory topic cue; they are not representative human semantic gold or
  actual voice/model answers. Old exposed references were not replayed.
  Lexical contract.
• Keyboard reading at f0353df: 286/286 related reader/detail tests,
  full TypeScript and isolated frontend build pass. Exact-head
  CI40790119879
  passed 4/4 jobs, including unsigned arm64 bundle/runtime checks.
  Native Chrome keyboard activation reproduced the original focus loss on
  generated data; the correction preserves pending/non-EOF/reset focus, blocks
  duplicate reads, transfers EOF focus and respects a user focus change.
  After CI, one retained original in the22-document snapshot passed six
  first/next/reset transitions at light1024x768 and dark1440x1000. Native focus,
  DOM SHA-256, byte ranges and EOF matched an independent raw-capture oracle;
  all40 state files preserved bytes/modes/paths after cleanup. This is source
  and isolated component proof, not installed Alden/native/GPU/physical-input or
  actual model-answer acceptance. The frozen reference baseline was not tuned
  or replayed. Contract.
• A separate source-only annotator froze16 questions and41 original-byte
  quotes from8 retained technical documents before lookup. At b87d009,
  indexed-lexical retrieval with5 entities/1200-character excerpts
  included the expected source first in10/16 and within five in16/16; all
  annotated support ranges fit the delivered excerpt in9/16. Applying the voice
  path's2-or-4 entity/400-character budgets delivered the expected source in
  13/16 and all support ranges in6/16. These are author-selected reference
  measurements, not representative semantic gold or actual model-answer
  accuracy. The graph UI's AND literal filter is a separate API. No query/answer
  tuning, embedding calls, product edits or native app startup occurred;40 state
  files and the frozen references remained unchanged. Evidence sufficiency,
  independent semantic/time/entity adjudication and actual answers remain open.
• The existing file-list entry now has actual retained-data component acceptance.
  In a user-authorized isolated Chrome for Testing context, the source
  settings markup and public controller selection binding displayed12 files.
  Mouse, Tab/Enter and Tab/Space selected the same admitted reader target.
  Across two themes and two desktop sizes,12 first/next/reset transitions
  matched independent original-byte hashes and ranges for that one document.
  Backend list loads, body failures and console warnings/errors were zero;
  all40 retained state files preserved their bytes/modes/paths after cleanup.
  This is a component display/selection test; controller.start, the other11
  original bodies, native app/installed readiness and full semantic gold remain
  separate. That earlier checkpoint used b87d009 and its4/4 CI.
• Full-original/scoped-query component at b87d009: 636/636 affected
  backend/packaging tests,10 new methods/124 subtests and93 independent native
  checks pass with stable source pins. The preserved pre-change typed fixture
  selected the wrong head document; the corrected host query selects its unique
  retained tail and persists exact proof plus hash-only request/query trace.
  Existing intent, original messages, permissions and zero-argument tool remain.
  Unknown grammar falls back unchanged; NFC expansion preserves the old input
  truncation order. Prior generated original gold passes96 search conditions,
  24 graph/direct conditions and448 hash/scope checks.108 fresh-process affinity
  tests and independent148+34+9 owner/resource controls pass; the permanent SQL
  authorizer remains unchanged and legacy API behavior is preserved.
  Exact-head CI40770362365 passed4/4 jobs, including unsigned arm64 bundle
  and runtime-resource checks. The same byte-verified22 historical original
  snapshots (182,711 bytes) pass12/12 tail searches,6/6 graph queries,6/6 strict
  affinity references and12/12 MCP searches across both strategies. Before index,
  the same12 queries had0 hits. A corrected typed HTTP mock completes two hops
  with one authorized read, original hash/proof and durable derivation trace.
  All29 retained data files stay byte-identical; initial coverage remains
  20 complete/2 partial. No new original acquisition, current original-disk
  permission, real decoder, installed app or normal release is claimed.
  Native review is separate from the held formal external audit.
• Current retained-original reader component at b87d009: two actual historical
  original snapshots (19,407 and28,078 bytes) across two themes and two desktop
  sizes pass24/24 first/next/reset DOM hash/range/EOF checks against a separately
  decoded raw-byte oracle. There are12 distinct layout cases and28 observations,
  with0 horizontal/source/console errors. All40 copied state files keep exact
  bytes and modes; source-only SQLite coordination timestamps are separate.
  Five native inline captures were inspected. Image-file export was rejected by
  browser URL policy and stopped; no alternate route was used. The overview
  samples12 source-file references and shows0 edges/visible caption labels;
  source-only entry affordance is under review, not accepted as full graph
  quality. This is source-component evidence, not native, installed, GPU,
  semantic-quality or real-decoder acceptance. The owned server/tab are closed.
• Retained-original window source at dc0afcf: 460/460 affected backend tests
  and 322/322 focused frontend tests pass, with unchanged current source pins.
  Full TypeScript and isolated Vite pass. Real generated collection/capture/API
  fixtures reconstruct exact UTF-8 bytes after original paths disappear, reject
  reflected permission/version/config/capture changes and preserve independent
  file IDs for equal bytes. Boundaries cover BOM/NFD/CRLF/emoji, strict cursor and
  schema checks, whole-original hashes and no parser/network fallback.
  Independent review reproduced numeric exponent overflow in strict capture JSON;
  the fixed decoder passes6 numeric and3 generated-reader closure cases. Its
  earlier MCP overflow probe stopped at the existing abort frame and was not
  retried; no wire-level closure is claimed for that probe.
  Generated real HTTP/API + reader rendering passes28/28 states across two themes
  and two sizes, including exact displayed hashes, EOF/partial separation,
  permission denial and tampered response hashes. Existing source coverage and
  excerpts are retained; errors clear the prior visible original. These are
  component/fixture checks, not installed Alden or private-original acceptance.
  CI40741979176 at
  dc0afcfba8573b5540efc83b9ded61e20ffe83b2 passed4/4 jobs, including unsigned
  arm64 bundle/runtime checks. After CI, current stdio MCP over a byte-verified
  copy of all29 retained-state files read22 actual original snapshots through
  24 windows: all182,711 bytes and individual input hashes match the independent
  captured originals. Both initially partial documents (19,407 and28,078 bytes)
  reconstruct exactly; their original parser receipts remain partial after EOF.
  Stale versions, unknown targets, invalid cursors and outside scopes withhold
  bodies. State hashes remain identical; guarded network/embedding/original
  fallback calls are zero and the owned MCP server exits0. This is retained
  snapshot acceptance, not fresh acquisition, current original-disk permission,
  live app state, installed reader or actual model inference acceptance.
• Original-file changes: 296/296 affected tests pass, including 53 parser and
  25 file-set integration methods; source hashes match before/after. Independent
  local review reproduced three defects, and independently reran their minimal
  closure probes after the fixes. All three are accepted within that scope.
  48/48 packaging/build-receipt tests pass. New modules are declared in the
  desktop resource maps and the prepared runtime includes their import closure.
  Focused Rust resource tests pass7/7. The whole-tree formatting check reports
  existing differences; the changed resource file has the same formatter diff
  as its c0b5d88 baseline. No broad reformat is included.
• Seven generated formats pass the actual registration → collection → retained
  capture → scoped FTS path. PDFKit and textutil execute in owned parser children;
  all seven originals retain exact bytes, including linked PPTX notes. Only the
  small Markdown and CSV bodies claim complete text coverage. Owned readonly MCP
  stdio returns seven search results, seven graph nodes and seven version-bound
  details; stale/outside requests reject, state hashes match and guarded
  network/embedding calls remain zero. These are generated originals, not private
  Spark corpus or installed-desktop acceptance.
• Removing either source-change or target-change publication guards in owned
  copied code makes its regression fail; unmodified controls pass. This is
  bounded mutation sensitivity, not an implementation-wide proof.
• Exact source-head CI at 058bad1d60aa78af6350a816251347ecc2f40a3a passed
  4/4 jobs: run40715508627,
  including the focused suites and unsigned arm64 resource/runtime checks.
  After CI,22 actual technical Markdown originals (182,711 bytes) passed
  collection/capture into task-owned private state,22 per-file FTS checks and22
  version-bound MCP source reads. All22 original-byte hashes match the retained
  historical index. Replaying the unchanged set preserves22 versions and the
  capture identity.20 bodies are complete;2 exceed the16,000-byte parser bound.
  Actual source files, the current Spark index and live app state were not
  modified. This is a finite technical subset, not full Spark discovery.
• Original-file coverage at 612e5c0: 421/421 affected backend tests and
  153/153 current reader tests pass; full TypeScript and isolated Vite pass.
  A bounded parser receipt binds the exact source, version, archive and delivered
  body. It reaches details, search/MCP provenance and the typed tool-role model
  input. Missing or invalid file coverage stays unknown; the existing response
  truncated flag still describes only the12,000-character excerpt. Reader
  notices precede long bodies and discard stale async hash results.
  Independent review reproduced a post-verification mutation gap; its25-case
  closure passes. Generated IAB component checks pass24/24 across two themes
  and two viewport sizes. Current read-only MCP over a byte-verified copy of the
  prior22 retained originals reports20complete/2partial with unchanged state,
  stale/outside rejection and zero guarded network/embedding calls. These are
  not fresh acquisition, installed-desktop or actual decoder acceptance.
  Exact source-head CI40732267381
  passed4/4 jobs, including unsigned arm64 package/runtime checks.
  Exact retained-original window source and post-CI acceptance follow below.
• Earlier unchanged trend aggregation at c0b5d88: 336/336 related tests pass;
  independent proof/date cases match. Contract and limits.
  Its CI run40687380021
  passed 4/4 jobs. Original-data A/B admission saw no existing WAL/SHM in45s:
  zero SQL opens and zero accepted pairs; no live performance gain is claimed.
• Earlier graph transport/layout suites and original-data component rendering
  retain their recorded revisions. Authorized isolated Chrome for Testing
  exercised pointer/keyboard selection, pin/reflow, bounded expansion,
  stop/resume and disposal over 1,984 nodes / 2,693 display groups. Actual hiding,
  native/GPU, physical Retina and fresh positive activity remain open.
  Failure contract,
  pin measurements.
• Current local candidate 0.3.220 / 7b4a57f includes the original-file
  features. One normal build and one canonical check pass; 2,135 tracked
  source files match the primary checkout by bytes, modes and Git objects.
  The app has 103 files /97 script resources /89 Python files. All96 source
  resources match current source; the one declared generated native voice
  library matches the fresh normal build and its pinned Swift sources. An
  initial parent comparator wrongly compared that generated library with an
  older ignored binary in the primary checkout; correcting only that binding
  required no source change or rebuild. App arm64 identity, strict/deep
  signatures, installed designated-requirement continuity and both bundled
  CLI version outputs pass. ZIP 22,562,559 bytes, SHA256
  5584830b40b499dd091614ba8f2d5b9593b19a95ce10bd34b298554f7efb3e19;
  extracted103-file bytes/modes/signatures match. The previous99-file candidate,
  installed0.3.21985-file app, current settings and protected service snapshots
  are preserved. Independent source/resource/ZIP verification passes with seven
  in-memory invalid controls; these are not product mutation proofs. This is
  Apple Development signing with linker ad-hoc CLIs;
  Developer ID, notarization, installed runtime and release remain unverified.

Alden remains OFF. Broader Spark original-file acquisition still requires exact
source-scope reconciliation. Broader file-format/full-original search coverage,
full source/final fixed27B and IQ3.3 inference, latest GBrain promotion, semantic gold, native input/GPU,
installation and release remain open. Supplemental native/Argon feedback does
not replace the reserved formal web-Astra → Gemini → reverse review; that stage
is still held. No local model/service adoption, global route/account/permission/
SDK/capacity mutation, local app startup, notarization or release is included. The full
original18, follow-ups, ST01–16 and Cursor objective remain active. This draft
source candidate remains externally unvalidated for release.

Async extraction reference fence (2026-11-01)

A source-extraction hash could finish after a node or selection reference changed and evaluate an introduced getter. The verifier and its separate DOM continuation now check own property descriptors before reading those live references, keeping stale extraction status unknown and original-body text hidden.

The repair includes two deterministic delayed-sibling regressions and 26 reference-mutation cases across the hash/DOM boundaries. Local validation: all 1,864 frontend tests, TypeScript and Vite build passed. Exact commit 5f4843f1b577cfeb3e88fa33a112058a6b9c0590 passed all four CI jobs. Existing draft/release holds remain; this is not installed-app, live-model, browser/GPU or deployment verification.