Project status and limits

The current capability map, scoped evidence and explicit non-claims.

Status: documentation preparation under ADR-010 W15. This index maps published operations to interfaces and evidence. It does not flip packet done or acceptance verified. Artifact identity must match before reusing a dated probe against a rebuilt binary.

Current verification boundary (9 November 2026)#

The historical counts below describe their named source and artifact cuts. The current health report records that the full profile failed at that audit cut and left runtime qualification incomplete. Its component passes do not renew the older packaged-service or whole-suite claims for current HEAD. Public policy verification checks planning and evidence identities; it does not execute those runtime probes. D01 remains partial until every published example is checked against the final integrated artifacts. Missing raw evidence remains unavailable; matching document hashes establish reviewed bytes only.

Landed RN-02a/RN-02b migrate service probes to protocol 2 with explicit receipt epochs and retained retry bodies; RN-03 maps InvalidProtocol to invalid_request. RN-04 repairs crash-child builds, with 537 abrupt-kill checks reported green by MAIN. RN-11b repairs CLI backup's protocol and wires the six short verbs listed below. These outcomes do not establish a passing full profile: RN-05 remains partial and RN-06 has not landed. Historical runtime counts remain tied to their named artifacts; D01 remains partial.

Evidence precedence (reconciled 12 October 2026)#

Read dated entries chronologically for the same named profile; the backlog evidence boundary does not treat later component checks as whole-profile acceptance.

The quarantine/pending statements below describe retained source cuts, not the latest result for those subsets; later entries preserve the failed diagnostics (counter/feed correction, route correction). Bounded PG3/WAL polling is not durable subscriptions or Application /v1/changes (feed qualification, MCP polling distinction); B19's authorized snapshot-to-durable-feed handover remains open in the backlog.

No upgrade follows for PF10/PF11 or whole-product acceptance (explicit non-claims, embedded limits), MCP catalog/context/expand or the five-answer Oracle walkthrough (MCP guide, B04), or actual UBI/native/import/reopen/restore/corpus acceptance (consumer prerequisites, inventory limits). The historical optional-hash receipt defect's repair status remains unresolved here: the formal refinement map explicitly does not reassess that repair, so later retry checks are not treated as FM04 closure.

Evidence gaps: .local/core-search-mesh-2026-10-04/A/ is absent in this checkout (12 October file-existence check), so raw logs and A/*.json receipts cited below could not be independently rechecked; this summary uses this file's dated records, not a new test run. The foundation packet's former PROJECT-INTERNAL/ target below is unresolved; no public replacement is established, and private archive content is not substituted (B42 reading boundary).

Legend for the Class column:

Local core/search/mesh candidate (4 October 2026)#

These source components have focused checks in the integration tree and are not part of the preserved qualified binary. Native batch groups admit at most eight requests/eight declared mutations/80 KiB encoded intents and publish fresh effects once. Independent fixtures verify exact receipts, rollback, mixed replay, selected rename failures and recovery fences. This does not qualify callback transaction queues or document/index hook parity.

Canonical values support exact mixed integer/finite-float comparison, binary UTF-8 string order, SQL three-valued truth and checked JSON integers. The ordered numeric key profile admits finite floats with magnitude below 2^64; current product integer schema/grammar stays separately bounded. Component checks do not imply all query consumers have been verified.

Security evidence covers trusted derived-key ownership, existing primitive vectors, published HKDF/X25519 examples and bounded TLS label, CertificateVerify input and framing components. Session key freshness remains the trusted caller's obligation; raw constructors cannot ensure uniqueness between owners. Separate five-fixture checks qualify the fixed whole-frame mutual handshake and its whole-record encryption wrapper with independent ciphertext/transcript recipes. Network fragmentation, key update, mesh transport and general application routes remain unqualified. Trusted borrowed credentials and key owners must remain stable and immutable until close/drain.

The current exact-i64 consumer composition passed scalar-consumer, schema/index and indexed-query checks without widening number admission. The separate query counter unit is quarantined after two unsuccessful repairs; its candidate and red logs remain conserved and its changes are excluded from integration.

Offline native backup streams a pinned immutable run into a checksummed artifact. Restore requires a trusted exclusive-recovery attestation and an absent destination; identity and receipts are preserved. Nine tiny fixtures cover replay, caps, corrupt or truncated input, allocation rollback, symlink/FIFO refusal and selected publication cuts, plus a retained capture while the source commits and compacts. A failed observation hook releases the pin before any artifact write. Original-writer fencing remains an external custody requirement. These selected cutpoints do not prove power-loss recovery or cancellation behavior.

Literal UTF-8 analysis uses ASCII separators/casefold while preserving non-ASCII bytes without normalization. Exact BM25 and vector ranking inspect complete supplied slices, compare the complete caller-provided root/policy tuple and retain it in results. Their bounded checks qualify arithmetic, ranking and input refusal. The text collection adapter additionally passed five bounded extraction fixtures, and its internal synchronous delivery gate passed three independent authority/root/expiry fixtures. It admits exact branch/collection scope before capture, owns copied corpus strings, and verifies the same authority plus actual physical root/context and fresh half-open expiry inside both gates immediately before callback. This profile grants whole collections, not row ACLs; no public wire endpoint or durable positional index is qualified. Vector source authorization remains unqualified.

Native branch drop/version history and dense retained-context suffix audit passed six bounded fixtures, including old pins, descendants, reopen and selected rename cuts. Existing native collection-feed composition passed ten checks. A separate direct native-feed oracle is quarantined after two unsuccessful repairs; its final result was five of six fixtures, with an unresolved boolean failure in retention coverage. A static IO model suggests the fixture cumulative-read assertion may be too small, but the exact failed line/count is unmeasured. The full C03 qualification remains blocked; no oracle was relaxed or third repair run. The separate three-voter/eight-entry chronological consensus model/checker passed six finite histories; no durable consensus runtime or real process cluster is qualified.

Matching logs are retained under .local/core-search-mesh-2026-10-04/A/; canonical package status and remaining gates are in the core/search/mesh registry and packet.

Current foundation and query subsets#

The local candidate uses TX3 intents, catalog3 definitions and strict LIM1 sidecars. It refuses obsolete backup1, outer consent1, pending1, headed ingest spools, patch wire spelling, dev-store, raw JSONRPC0 and importer progress1. Existing records are not converted or rewritten. Current journal/full-versus- compacted and backup/full-versus-checkpoint forms, inner mandate1 and consent2/3, plus pending2..8 domain kinds retain their distinct contracts. The former foundation packet reference, ../../PROJECT-INTERNAL/MANAGEMENT/SPRINT-ORCHESTRATION-FOUNDATION-SANITY-2026-10-02.md, is an unresolved evidence-link gap: matching checks and remaining decisions cannot be rechecked through that target (12 October file-existence check). As of 9 November 2026, version 0.0.44 retains these compatibility changes; this is no performance result or acceptance of deferred product requirements.

The current bounded additions have separate guides: indexed reads, SELECT, SUM, SQL SUM, scalar membership, grouped COUNT, SQL grouped COUNT, inner equality joins, path updates, streaming recovery, image recovery, RAM image transfer, RAM maintenance, RAM branches, collection feeds and browser numeric traces. They qualify stated local subsets; broader SQL, durable profiles, paging and distributed security remain open.

Entry points#

Data and analytics#

Recovery and faults#

Explicit non-claims#

• No production, hosted, real-money, notarization or clean-machine rebuild claim.

• No accepted cold/warm/token/latency Oracle envelope.

• No verified O01/V09 five-answer sole-owner walkthrough.

• No performance qualification (PF10/PF11) under this lane.

• P27/P28/P29 done/verified remain Eris decisions.

See service CLI, MCP local, documentation round and docs evidence.

Selected encrypted native-run profile (5 October 2026)#

The integration tree supports an experimental trusted-provider encrypted immutable native-run profile through the existing writer, publication, bounded cache and recovery owners. Nine bounded native fixtures and six independent AEAD fixtures passed. The current source also passed unchanged clear paging (25) and affected lifecycle (6), backup (9), collection-feed (10), budget (3) and maintained selective-index Application (5) checks. The direct native-feed candidate remains quarantined and does not gain qualification from these checks.

Fresh OS object identities derive per-object keys; a monotonic per-writer encryption counter covers both writes of page zero. Counted physical records are 4,120 bytes plus a 128-byte object header. Provider secrets and owned decrypted cache/scratch buffers have scoped disposal. Metadata and artifact headers remain visible. The provider must remain trusted and immutable for the database lifetime; persistent key-vault custody, automatic revocation, WAL/checkpoint/outbox encryption, historical clear-file migration and rollback protection are absent. The protected shipped binary is unchanged.

The empty retained-suffix audit fix anchors only the selected nonzero current metadata when no retained requests exist. It proves no retired request or context history. Compiled prologue samples are frame lower bounds; they do not establish all-path total-stack bounds or qualify the existing 448 KiB allowance.

Declared scalar, vector and consensus model profiles#

Canonical ordering covers missing, null, booleans, signed/unsigned integers, finite binary64 comparison and binary UTF-8 strings, with SQL unknown separate from document absence. The 143-byte magnitude payload (145-byte complete numeric key) admits integers and finite floating magnitudes below 2^64; unsupported decimal/object/array ordered keys refuse explicitly. Checked JSON integer parsing and current integer-only plan/index consumers share this owner. Fresh checks passed scalar 8, consumer 2, schema 15 and indexed query 3; wider schema/query numeric grammar is unsupported.

Exhaustive vector search implements dot, cosine and squared L2 for complete bounded collections: at most 64 live documents, 256 dimensions and 16 results. Stored and query numbers use exact lexical binary32 admission; calculations use fixed-order binary64. Actual approved collection extraction passed 5 checks, internal delivery 4 and metric/reference fixtures 4. The one-time trusted space profile is copied and compared before capture or query decoding. Results retain exact scope/profile, source revisions and physical root binding. Final authority, root/context and callback-time expiry are checked inside the existing gates. This is an internal trusted embedding profile, without a durable vector-space catalog, row ACL, ANN or public wire endpoint.

The fixed three-voter consensus model and hostile-history checker passed 6 fresh finite fixture groups. It maps terms, votes, log changes and responses to planned canonical persistence transitions; it does not implement a durable consensus runtime or qualify real replication. These declared bounded profiles do not complete the 45-package program or bypass the quarantined counter/feed dependencies.

Current bounded recovery and literal analyzer checks#

Native recovery now has ten matching checks, including an actual document restored at revision2/body string30 and immutable request replays retaining sequence1/2 and the selected root. Identity-preserving recovery requires an externally fenced original writer; these checks do not prove global fencing or power-loss recovery. The native batch-group profile admits batch requests only; Application callbacks and maintained-index lowering remain unsupported.

The literal UTF-8/ASCII analyzer now has five checks including safe original-offset HTML-text highlighting, capped at8192 input bytes,256 spans and52480 output bytes. Its provenance is RFC3629 sections3/4 and Zig0.16 std.unicode source hash6b7243b43bb1467b48d3027bcb379078f9ca077086bfe6f887d04de6259c2934. It does not implement UAX word-break or NFC. Matching BM254/4, text-source5/5 and internal text-delivery3/3 passed. Native encrypted storage now has ten checks including three exact low-provider refusals with terminal key disposal/no accepted manifest/adoption; earlier nine-check evidence remains preserved. These counts qualify the declared components only.

Selected primitive ownership and local key updates#

The declared primitive profile has matching checks for full pinned ML-KEM encapsulation/decapsulation/key-validation and ML-DSA verification/malformed inputs (8), derived synthetic-session ownership/disposal/allocation cases (11), pinned derived vector (1), TLS label/framing literals (4) and schedule literals (3). Current KeyUpdate composition rechecked these selected owners. This qualifies the named vectors and ownership profile, not ACVP/FIPS certification, side-channel resistance or a protocol audit. Zig standard-library provenance remains distinct from first-party runtime code.

The immutable-pin local TLS component now supports published directional KeyUpdate0/1, pending and crossed responses, old-key update/new-key sequence-zero transitions and one-shot application-secret transfer. Five new literal groups and matching unchanged session5, handshake5, record5 and schedule3 passed. One legacy application-only typed-error compatibility repair preserves its existing InvalidInner refusal; the failed4/5 composition log and original source remain retained. Dynamic current-peer trust, local service revocation and public transport delivery remain unqualified.

Bounded rebuilt positional search and current pin custody (5 October)#

The integration candidate now builds complete owned single-field posting snapshots from the same authorized canonical document capture. It retains revisions, original offsets, positions and authorized corpus statistics; exhaustive BM25 and ordered phrase eligibility preserve deterministic ties. The profile admits 64 live documents, 256 global terms, 4096 positions and 65536 term bytes under the existing counted memory owner. Snapshots rebuild after canonical updates, deletes and reopen; they are not persisted posting indexes or accelerated pruning. Explicit delivery opt-in retains the same authority, exact-root and callback-time expiry checks. Matching posting/reference checks passed 4/4 each, delivery 5/5, vector delivery 4/4 and text source 5/5.

A Session can bind one stable owner-provisioned current peer-pin custody owner before handshake. Its serialized pin check covers guarded application and KeyUpdate operations; revoke, replace and fixed-owner substitution refusal preserve output canaries and retire Session secrets. Six bounded custody checks and matching KeyUpdate/session/handshake/raw-key checks passed. This is local trusted pin custody, with no remote registry, network endpoint or universal memory-erasure claim. Peer-input error normalization remains under separate review.

Published local TLS ingress profile qualification (5 October)#

The bounded TLS Session adds explicit local peer-ingress methods with a finite classifier for malformed frame, profile, keyshare, signature, Finished, authenticated inner-type and KeyUpdate failures. Classified peer failures return InvalidPeer after terminal secret disposal; owner, capacity, cancellation, current pin trust and unknown provider errors retain their typed meaning. Peer public-key validation and local stored-secret failures are distinguished by the owned handshake phase. Current matching Session checks passed 8/8, KeyUpdate 5/5 and pin custody 6/6. Together with the retained independently generated transcript/AEAD literals, this qualifies the declared local published-profile component. It provides no socket/alert implementation, remote registry, interoperability certificate or whole-product security audit.

The broader S06 route audit remains partial. Fresh collection feed 10/10, receipt policy 1/1, service 4/4, authority 5/5 and hostile 5/5 passed. Older WAL Application fixtures compile after reviewed namespace adaptations but fail at runtime: Application 0/11 after its final bounded repair, graph 0/9, work 1/8 and indexed 0/4. Their error owner remains unresolved; the indexed response reports storage_unavailable. All failed evidence is retained, no backend conversion or weakened oracle was accepted, and no third Application repair is authorized.

Selected direct encrypted Journal qualification (5 October)#

The explicit normal version-6 Journal profile passed 9/9 independent checks, with matching object-envelope 5/5, existing native crypto 6/6, encrypted native storage 10/10 and native budget/lifetime 3/3 checks. Replay authenticates and validates the complete private physical image before exposing canonical payloads and retains physical offsets. Tests cover trusted identity admission, a reauthenticated inconsistent inner catalog identity, checksum/sequence/counter/padding/length failures, torn suffixes, provider and allocation refusal, physical group limits and uncertain sync outcomes. Runtime was 328 ms/2 MiB; the guard peaked at 958 MiB. The caller allocator and physical log limit govern direct Journal operations.

Each attempted frame receives a fresh OS object identity. A group writes ordered frames and one sync; recovery may accept a complete valid prefix containing fewer frames than the attempted group. This is not atomic logical group replay or power-loss proof. Encrypted Database open, checkpoint suffix, capture, copy, compact, same-handle repair, backup and restore remain unsupported. This evidence does not qualify a key vault, rollback anchor or encryption of every durable file.

Selected direct encrypted checkpoint qualification (5 October)#

Direct checkpoint publication and loading passed 9/9 independent checks with matching journal 9/9, encrypted native storage 10/10 and budget/lifetime 3/3 composition. The explicit MF2 manifest binds ciphertext length and hash through the existing CURRENT publication owner. Complete image authentication precedes canonical DTO exposure; the decoded store, publication generation, sequence, physical journal offset and prefix digest must match the selected manifest. Independently serialized fixtures distinguish generation from sequence. Actual canonical state preserves document revision 2 and both complete immutable receipts. Tests cover repaired public hashes reaching AEAD failure, current identity admission, allocation rollback and selected publication sync cuts.

Encrypted publication advances one checked generation and prunes one obsolete managed generation pair. Earlier failure orphans may remain. The direct component trusts the caller's journal boundary and prefix; synthetic fixture prefix bytes do not prove journal custody. Runtime was 399 ms with 64 MiB test RSS, and the guard peaked at 968 MiB. This uses a caller allocator and bounded image limits. Private physical/page/key buffers are wiped; the existing logical DTO owner does not provide a universal memory-erasure guarantee. Database, suffix, capture, backup and restore encryption remain unsupported, and no key-vault, antirollback or all-file encryption claim follows.

Selected encrypted Database runtime (5 October 2026)#

The explicit Database.openEncrypted profile passed nine bounded checks on Journal 1f46e36f, Database 638dd8ca, restore admission guard 5c6819f7 and independent fixture 9f006a66. The first run reported 441 ms and 4 MiB test RSS; sampled guard RSS was 1081 MiB. It admits normal encrypted journals without CURRENT and single batch requests through the existing Engine. Full authenticated replay also refuses unsupported nonbatch history. Any CURRENT entry refuses under the journal owner lock before journal creation, provider access or sync. Generic Database.open continues to refuse encrypted options.

Immutable request replay preserves complete receipts and ciphertext bytes without another provider call. Existing store-limit metadata is mandatory on reopen and remains plaintext; it is not silently repaired. Wrong trusted identity/key, provider failure, allocation failure, physical log limits, torn frames and authentication failures have independent refusal witnesses. A selected sync-unknown outcome fences new commits while the previous published snapshot remains readable; reopen authenticates the durable suffix before accepting its effects. Snapshots may outlive Database close under their existing state owner and caller allocator. The provider must remain stable through Database operations and close.

Hooks, publication observers, groups, nonbatch commands, checkpoint publication/loading through this runtime, all captured histories/feeds, copying, backup and restore remain unsupported before their effects. Restore's encrypted-options guard precedes artifact access and staging. This is a trusted typed engine profile, not an Application endpoint, service-grant acceptance, encrypted checkpoint runtime, key vault, rollback protection or all-file encryption. Private replay and temporary encoded-request buffers are wiped; retained Engine and decoded DTO allocations use existing disposal and carry no universal erasure claim.

Matching affected checks passed: direct Journal 9/9 (386 ms/2 MiB), direct checkpoint 9/9 (355 ms/64 MiB), native backup 10/10 (350 ms/3 MiB), encrypted native 10/10 (383 ms/4 MiB), and native-budget/lifetime 3/3 (297 ms/2 MiB). These caller-allocator/runtime measurements do not qualify a total stack bound or alter native memory limits. S04 remains partial; I02/C03 quarantines and unresolved S06 Application failures remain retained.

Selected encrypted checkpoint Database runtime (5 October 2026)#

The explicit Database.openEncryptedCheckpoint profile passed nine independent bounded checks, followed by matching no-checkpoint Database 9/9, direct Journal 9/9, direct checkpoint 9/9, native backup 10/10, encrypted native 10/10 and native-budget/lifetime 3/3 checks. The new test ran in 352 ms with 6 MiB test RSS; its unchanged build guard peaked at 1052 MiB. These measurements use a caller allocator and do not qualify a total stack bound or change native limits.

The constructor requires an existing encrypted journal and CURRENT. It authenticates the complete journal, validates batch grammar across both prefix and suffix, verifies the actual ciphertext prefix digest and exact physical checkpoint boundary/sequence, and imports through the existing Engine before applying the suffix. Checkpoint publication captures a coherent Engine freeze and physical journal boundary under the existing writer gates, then uses the same canonical image and CURRENT publication owner. Publication generation remains distinct from request sequence. Tests directly recover document revision 2 and both full immutable receipts, exercise valid prefix/suffix replay and selected publication failures, and refuse repaired authenticated boundary/profile inconsistencies.

CURRENT, manifest, image and store-limit reads check named-file metadata before a potentially blocking read-only open. No-writer FIFO fixtures refuse with the existing typed errors before the relevant provider access; descriptor checks remain in place. This prevents the tested named-file hazard under the owned directory contract and does not establish safety against a foreign filesystem mutator racing metadata checks. Synthetic provider cancellation and bounded allocation failures preserve typed local errors; they do not prove operating-system cancellation.

The accepted checkpoint state comes from a trusted producer and key provider. Ciphertext prefix binding is not an independent proof of checkpoint state equivalence or antirollback. The profile remains single-batch, branch 1, context 1/1 and receipt epoch 1; public captured histories/feeds, encrypted backup/restore and compaction remain unsupported. The metadata sidecar remains plaintext. Private replay and encoded request/image buffers are wiped; retained Engine and logical DTO allocations carry no universal erasure claim. S04 remains partial, and the I02/C03 quarantines and unresolved S06 failures remain unchanged.

Renewed corrective evidence (5 October)#

Human continuation reopened the retained counter, direct-feed and Application corrective units; the earlier failed candidates and logs remain historical evidence. The coherent optional SELECT counter profile now passes its original two checks with checked cumulative observations, construction/copy accounting, disabled-output parity and unchanged refusal/overflow assertions. This is a component result: the broader I02 cold-admission/selective baseline remains pending.

The original six direct native-feed checks now pass. The measured prior retention failure read 3,211,264 bytes; an independent per-stage no-cache bound is 4,190,208 bytes. Only that first owner's fixture estimate became 5 MiB; its 2 MiB write estimate, all other fixture estimates and the production 256 MiB read/32 MiB write caps remain unchanged. Matching lifecycle and collection-feed checks passed 6 and 10. This qualifies the declared ordinary native lifecycle profile, including full-frame feed/replay, retained suffix/resnapshot and reader-pin retirement fences. Public projection, replica obligations and unsupported operation profiles gain no qualification.

The positive Application protocol helpers now use the canonical protocol. Graph checks passed 9; the other three route roots still require reviewed public identity-error mapping, receipt-shape compatibility and retained-proof cleanup corrections. S06 remains partial while these and the named exposed-route matrix checks are pending. S04 also remains partial; no encrypted backup/restore/compaction or all-file protection is implied.

Matching exposed-route correction evidence (5 October)#

The four Application route roots passed 11, 9, 8 and 4 checks without allocator leaks. Positive envelopes use the canonical protocol and a once-captured receipt epoch for new intents; saved retries remain immutable. Public wrong-store refusal remains the declared stale_session, while the internal error stays typed WrongStore. Receipt comparisons include the immutable epoch, and callers release retained index proofs before closing their database. Bridge checks passed 5, including the historical v1 fixture and current wrong-store/session refusals.

Matching current checks passed authority 5, hostile 5, service 4, text delivery 5, vector delivery 4, native collection feed 10 and feed policy 1. The exposed intent, mandate, wallet and proposal routes passed 2, 1, 1 and 1. The reviewed route inventory distinguishes the common owner gate, operation-specific control exceptions, index-proof custody and final current-root/authority search admission. Immutable bootstrap credentials are not a mutable peer or general-reader grant registry; direct trusted Engine APIs do not acquire universal ACL enforcement from these checks. SQL and nested-query compositions also passed 3 each after current caller metadata and retained-proof disposal corrections. Historical failures and their diagnostics remain conserved.

These checks qualify the declared exposed owner/control and bounded approved-collection search profiles. General mutable-reader routes, wider engine authorization, public transport and heavy server acceptance remain outside that qualification. I02 remains partial until its declared bounded cold-admission/selective and posting-work baseline passes; S04 remains partial and C04 is still at explicit paged-catalog/chunk contract gates.

Bounded work baseline (5 October continuation)#

The I02 baseline passed 3 checks, with matching text 4, postings 4, original work 2 and search-delivery 5 checks. A 16-document native collection measured cold reopen admission separately from first, repeated and observation-disabled selective queries. Both observed queries visited, decoded, matched and constructed one row, copied 424 bytes and constructed 211 result bytes; the repeated query performed no admission audit. Disabled observations preserved exact output bytes. The owned posting build constructed 18 postings, and its query observed 16 documents, 32 frequency entries and one scored candidate.

Initial creation and seven subsequent selected durable publications wrote 2,316,288 provider bytes. Cold admission read 3,002,624 bytes; the first query added 1,593,344 and the repeated query added 94,208. Provider dispatches and accepted publications are distinct from OS syscalls and sync dispatches. Single elapsed samples are uncertain, and logical cold reopen does not control the OS cache or establish sustained performance or observation overhead. Existing 256 MiB read, 32 MiB write and counted 2 MiB native working limits remain unchanged. This qualifies the declared bounded selective and posting-work baseline, with no universal instrumentation or heavy server claim.

Journal ciphertext prefix capture (5 October continuation)#

Nine named Journal-only encrypted prefix checks passed (413 ms/2 MiB runtime), with matching direct Journal, checkpoint, no-CURRENT Database and checkpoint-aware Database checks passing nine each. Capture validates each private object and its canonical frame, copies original ciphertext, compares its reread hash before advancing validation, and accepts the full copy only after the frozen identity/sequence matches. Alias and output-end admission precede key-provider and write effects. Frozen copies may exclude later appended frames. Stored corruption and local provider, allocation and destination failures retain distinct custody behavior; a partial destination has no accepted copy result.

The read-byte model is at most twice the prefix length plus 128 bytes per object, with writes bounded by the prefix length; this is not a syscall measurement. Scratch belongs to the caller allocator, with one bounded physical object and a 64 KiB copy chunk. The largest fixture object was 70,168 bytes; the supported object maximum is 4,223,128 bytes. These facts do not qualify a total 2 MiB native memory bound. Cancellation is a synthetic key-provider refusal, and selected reread mutation is not a foreign-filesystem race guarantee. Database public capture, encrypted backup/restore and compaction remain unsupported; S04 remains partial. The first compiler failure and its narrow error-consumption repair are retained.

Named Database ciphertext copy (5 October continuation)#

Nine explicit encrypted Database log-copy checks passed (647 ms/5 MiB runtime), with matching no-CURRENT Database, checkpoint-aware Database, Journal prefix and direct Journal checks passing nine each. The existing writer/publication owners freeze the selected prefix and require the Engine and Journal sequences to agree. Copy then uses the qualified Journal ciphertext path. A selected key-provider callback appended revision 2 while the frozen revision-1 prefix was copied; both full immutable receipts, exact ciphertext hashes and document revisions were checked.

Admission rejects aliases and invalid output bounds before provider/write effects. Unknown commits and a sequence mismatch fence new capture and commits while preserving the prior published snapshot. Stored authentication failure poisons current capture custody; provider refusal, synthetic cancellation, allocation failure and a partial destination write retain distinct local errors. A partial destination receives no accepted copy result. Clear copy behavior and native/wrong-profile guards retain their declared behavior.

This is canonical ciphertext-prefix attestation under the trusted batch-profile constructor and stable key-provider premise. It does not independently establish Engine-state equivalence or grant restore admission. Callers keep the unmoving Database, provider and files alive and drain copy before closing; concurrent close is not qualified. Journal object scratch and the read-byte model remain caller-owned, without a total 2 MiB native-memory claim. Public plaintext captures, history/feed, encrypted backup/restore and compaction remain unsupported. The plaintext store-limits sidecar is explicit metadata; S04 remains partial.

Exact block-max ranking and bounded boosts (5 October continuation)#

Six component checks passed (541 ms/6 MiB test RSS), followed by seven search-delivery checks (381 ms/7 MiB). Matching reference text, postings, bounded work baseline and vector delivery passed 4/4, 4/4, 3/3 and 4/4. The snapshot and acceleration use the same captured binding, live statistics and counted allocator. Delivery remains explicitly selected and retains the final authority, root and fresh-clock checks. The original reference and null-mode paths retain their behavior.

Bounds use maxima of the exact finite, nonnegative per-document term contributions. Bounds accumulate in the same deduplicated lexical order as the reference scorer; monotone floating-point addition preserves domination. Global WAND pivots use global maxima, and block skipping uses a common document-rank interval. Skipping requires a bound strictly below the current threshold; equal bounds preserve identifier ties. Phrase queries and uncertain or nonfinite metadata use the complete exhaustive path. Results were compared bit for bit with that reference across unsorted identifiers, ties, actual 7/8/15/16-document boundaries, extreme 256-token frequency, disjoint term blocks, changed corpus statistics, tombstones and reopen.

The admitted boost is a uniform query multiplier of 1, 2 or 4. Validation precedes query parsing and capture. Raw ranking and bounds remain unchanged; exact binary scaling applies only to the complete fixed result. Positive normal finite score guards refuse unsupported scaling. Independent exponent-bit checks verify boosted score bits, identifiers, revisions and unchanged skip observations. Unsupported factors and missing explicit posting mode refuse before source access. Per-term, document, field and arbitrary boosts remain unsupported.

A 24-document independent fixture returns the literal winning identifier while observing eight candidates, eight scored documents, ten block checks, two skipped blocks and a cursor distance of 48 entries. Cursor distance counts list entries, not comparisons or all work. Optional counters reset at each call and use checked arithmetic; disabled observation preserves results. This bounded snapshot profile does not qualify durable maintained posting indexes, unrestricted text language, sustained performance, observation overhead or a speed improvement. Existing storage, query and compilation limits remain unchanged; historical compiler failures and preimages are retained.

Authorized dictionary subset (5 October continuation)#

An immutable authorized posting snapshot provides complete single-token prefix expansion and ordinary Unicode-scalar edit-distance suggestions at distance one or two. Binding admission precedes parsing; raw query bytes are limited to 8192, dictionary terms and query to 64 scalars, dictionary size to 256, output to 32 borrowed terms and DP work to 65536 checked cells. Exceeding a bound refuses the whole call. The existing analyzer folds ASCII case; no NFC or full Unicode word-breaking claim is made.

Dictionary checks passed 6/6 (385 ms, 3 MiB test RSS; 466 MiB sampled guard RSS), with matching reference 4/4, postings 4/4, block-max 6/6, bounded baseline 3/3, authorized text source 5/5 and delivery 7/7. The counted native fixture used 303104 read bytes, 209664 write bytes and an 808901-byte requested-memory peak under the unchanged default 1 MiB owner; reopen used 545024 read bytes and no writes. A denied malformed neighbor changed the root without changing authorized vocabulary. Historical compiler and missing-directory fixture failures remain preserved. R04 remains partial: snippets, facets, ranked prefix composition and new Gate delivery semantics are not qualified.

Direct clear paged catalog subset (5 October continuation)#

The explicit paged NM2 profile stores fixed branch facts under the same native selected root and publication owner, with a fixed 256-byte summary and 64-byte catalog rows. The default fixed NM1 profile is unchanged. This first direct profile admits clear batch, fork and drop operations at context and retention floors one; encryption, retirement, compaction, recovered reclamation, generic Database adapters, backup and restore remain refused. Branch count is bounded independently of full nonzero 64-bit branch IDs.

Full staged and reopened audit validates complete history shape before canonical history reads, exact namespace coverage, branch counts and record totals, each branch timeline and the parent state before child creation. A later parent deletion does not invalidate an earlier valid fork. Checked upper resource quotes precede reclamation and candidate writes; complete staged audit precedes CURRENT selection. Catalog rows are copied before nested reads, and independent cursors retain bounded scratch ownership. The named catalog pager seam admits twenty existing references plus one catalog row; ordinary pager entry points retain twenty.

Nine independent checks passed (644 ms, 4 MiB test RSS; 718 MiB sampled guard RSS). One live seventeen-branch creation owner reserved 188321792 read bytes and 7442688 write bytes, measured 93474816/7246080 provider bytes and a 687612-byte requested-memory peak. A separately drained cold owner reserved 155980032 read bytes, measured 12861696 read bytes and a 605044-byte peak. The twenty-one-reference witness peaked at 754056 bytes. All use the unchanged default 1 MiB counted memory owner and 256 MiB read/32 MiB write caps; separate owner totals are not aggregated into a single budget claim.

Matching paged storage 25/25, encrypted native 10/10, native backup 10/10, lifecycle 6/6, selective native Application 5/5, direct feed 6/6, collection feed 10/10, timeline conservation 2/2 and state-sequence model 2/2 passed. Existing proof pins, immutable unknown-outcome retries and full receipts remain explicit. Owned CURRENT, selected-run and reclaim-run named pipes refuse before blocking opens; no independently racing filesystem-mutator protection is claimed. Historical stale-envelope and compiler-export fixture failures remain conserved. C04 remains partial until chunk custody, incremental successor proof and profile propagation qualify; this subset is not a general scalability or total-stack certificate.

5 October: bounded snippets and encrypted compact-header leaf. Nine snippet checks passed with the same authorized original Corpus, scalar-safe windows, complete phrase matches, escaped output, whole-call limits and initialized allocation rollback. Matching dictionary/reference/posting/block-max/baseline/source/delivery checks passed 6/4/4/6/3/5/7. The snippet result retains the copied binding and revision; caller-supplied original text remains a trusted Corpus input. This does not qualify facets, fuzzy snippets or new service delivery.

Nine encrypted HEADER7 leaf checks passed after one fixture-only correction supplied the required clear Journal limits. A fresh authenticated 4,248-byte header retains the original physical base offset while replay and append use local physical positions. Independent ciphertext literals, unchanged suffix bytes, copied-file reopening, typed refusals, exhaustion, provider cancellation and allocation rollback are covered. Four matching Journal/cipher-capture/checkpoint-runtime/database-capture roots passed nine checks each. This qualifies encoding and direct Journal custody only; offline compaction publication, checkpoint adoption of compacted files and whole S04 remain unqualified. All failed evidence is retained; G12 and storage caps are unchanged.

Approved string facets (5 October bounded direct profile)#

The existing approved full-collection walker supports direct string facets under the same captured source binding and allocator. Limits are 32 buckets, 160 decoded bytes per value and 5,120 owned string bytes; callers may only lower them. Missing and null fields share a missing count; empty strings form a normal bucket. Counts cover every admitted live envelope, sorted by exact decoded string bytes. Unsupported field types or exhausted limits refuse the whole call.

Nine source-bound checks passed (417 ms, 9 MiB test MaxRSS; guarded build peak 701 MiB). Matching source 5, delivery 7, baseline 3, snippet 9 and dictionary 6 checks passed. The native fixture observed a 32,768-byte scan delta, 139,264 total read bytes, 102,912 write bytes and 808,751 requested-allocation peak under the unchanged 1 MiB owner cap. Deterministic allocation failures cover initialized buckets through final ownership transfer. This direct trusted captured-source API adds no filters, numeric or multivalue facets, top-K approximation, new Gate admission or whole-R04 completion.

5 October: chunk values, offline encrypted compaction and socket transport#

The direct clear NM2 chunk-value profile passed nine checks with matching catalog 9/9, paged-storage 25/25, native-feed 6/6, state-sequences 2/2 and encrypted-native 10/10. Values are bounded at 32 KiB and stored by one atomic five-PUT batch: a reference and four stable slots, with unused slots present and empty. Reads verify all slot lengths and the full value hash before the borrowed callback. The selected view supplies the same allocator and a full-root-bound audited profile marker; the marker is trusted internal metadata, not a sandbox capability. One owner reserved 12,066,816 read bytes and 906,240 write bytes, observed 5,042,176/709,632 provider bytes and 721,236 requested-allocation peak. The corrected fixture model includes all nine diagnostics on the final root; global 1 MiB/256 MiB/32 MiB caps are unchanged. Logical deletion, encrypted chunks, generic adapters and incremental successor proof remain outside this qualification. Evidence: A/C04-chunk-qualified-subset.json; whole C04 remains partial.

The named offline encrypted compactor passed 9/9 with four matching encrypted roots passing nine each, plus clear offline-compaction 6/6. It validates complete source checkpoint recovery before target effects, preserves checkpoint and suffix ciphertext, and publishes a fresh HEADER7 through the existing offline publication owner. Clear WAL restore preserves captured data and sequence but creates a fresh store/retry namespace; native C02 identity-preserving restore is a separate fenced profile. Original-store retries, old-store refusal, one fresh restored intent, full receipt and reopen parity are covered. Evidence: A/S04-offline-compact-qualified-subset.json. Compacted Database runtime adoption, standalone original-prefix authenticity and antirollback remain unqualified; whole S04 is partial. Earlier OOM, compiler and restore-oracle failures are preserved.

The bounded socket/session adapter passed eight checks, matching Session 8/8, handshake 5/5 and key-update 5/5, and an actual two-peer localhost exchange on macOS. The generic baseline Linux probe then passed a Mint-local exchange. One Mac-to-Mint exchange sent eight synthetic records each direction with independent crossed hashes and the same 32-byte transcript, finishing and draining both hosts in 8.344 seconds. Both test probe listeners stayed on loopback; installed OpenSSH provided development forwarding while RURAL Session mutual pins authenticated the inner byte stream. Mint peer and watcher quotas together stayed within 150% CPU, 4 GiB memory and 128 tasks; macOS peer plus tunnel RSS was sampled. Evidence: A/M02-two-host-qualified-subset.json. This qualifies positive transport correctness only, without a durable inbox, database replication, Authority grant, product mesh, post-quantum interoperability or performance claim. All process/controller refusal evidence and existing guarded-build limits remain retained.