AOCI-CODE v0.1.0-rc4

AOCI-CODE v0.1.0-rc4 is built from 42928c342d76f9273d2a672420631ab1c4886e37.

• Add a narrow openGauss 6.0.5 LTS A/PG Schema Evidence collector, backed by a
  reproducibly patched official Connector v1.0.8, strict remote verify-full
  TLS, fail-closed unsupported catalog handling, and disposable real-engine
  acceptance without changing Evidence v1 or the nine-tool MCP surface.
• Expand the fixed general-purpose Code and Database starter A/B dictionaries,
  make all C importance digits from 1 through 9 available, retain optional D
  grammar and the existing E scale, and use EG7T as the starter Code example.
• Reserve starter G for genuinely cross-domain objects and Z for understood
  objects that fit no named category; evidence gaps do not become Z or S
  constraints. The new defaults affect only fresh initialization and do not
  migrate or retag repositories with an existing formal Meta.
• Report exact Code candidate binding mismatches and distinguish
  code_plan.batch_id from the cross-domain authoring_batch.batch_identity
  without changing the nine-tool MCP surface or request Schema; actual source
  drift remains a stopped replan condition instead of a copied-field repair.
• Expose the aggregate Check command in an authoring-required Volumes Guide and
  close the final successful batch through Verify, Aggregate Check, and Guide
  while preserving intermediate-batch and Legacy Entries Stage behavior.
• Return the same Verify, Aggregate Check, and Guide closure directly from a
  successful final Volumes Apply, while leaving paged, Legacy, and Cognition
  Optimization actions unchanged.
• Advance an already-managed Root fingerprint in the same Database Cognition
  Bootstrap Baseline postimage as the Root descriptor update, and narrowly
  reconcile the canonical historical state left by earlier Bootstrap versions.
• Add a session cognition line to aoci_search, aoci_get_entries, and
  aoci_header, and an optional two-question cognition probe on a
  check_only Overview, so a Host can tell whether it still holds the
  delivered Whole-Index instead of re-requesting it. The line is computed from
  session facts with no repository scan, receipt identity drift never raises
  it, and the probe measures recall without advancing the refresh generation.
• Grade the Whole-Index Attestation as assimilation rather than verbatim
  recall: the Challenge passes at 80 percent or better with at most one object
  identity miss, core F is judged by normalized token similarity that splits
  Han, Kana, and Hangul into character bigrams so every Locale meets the same
  floor, and object identity and Tag stay exact. fail is reserved for a
  foreign envelope or no correct ordinal; every other shortfall records
  partial, so an honest complete-coverage claim can no longer grade below
  the same answers submitted with a hedged one.
• Accept host_delivery_confirmation and model_cognition_attestation in one
  call or in separate calls in either order. Both halves bind the same
  delivered body, so the session remembers each half per body and grades the
  merged evidence; an explicitly carried half supersedes the remembered one,
  and a fresh complete delivery resets that memory.
• Size one authoring batch for the Host tool-result window. The team key
  code_cognition_batch_entries (machine default 20, wire ceiling 200) sets
  how many candidates one Maintain asks the model to author inline, and
  Maintain keeps per-item governance enumerations to a leading sample with
  complete counts under governance.list_truncation and sets.review_total.
  Candidates, plans, and receipts stay complete, aoci verify and aoci check
  still list every item, and the Database Evidence byte gate defaults to
  64 KiB.
• Return structured repair findings instead of an internal error when an
  enforce-mode cognition budget rejects an Entry: every violated field carries
  candidate_index, field, actual_tokens, and max_tokens, whole-index
  excess and violations outside the batch stay a batch-level stop, and a CJK
  S field that crosses its token band is now a locatable repair.
• Accept model-authored R exactly as written: AOCI never checks one Entry's
  relations against another Entry, so a relation whose target is missing,
  unmanaged, scheduled for a later batch, or ambiguous by bare name is
  persisted unchanged and produces no Finding. aoci_update_entry no longer
  returns impact_relation_unresolved, impact_relation_ambiguous, or
  impact_relation_invalid; per-Entry FRAS structure, the tag dictionary, the
  C-driven S quota, source binding, and the projected budget stay enforced,
  and a newly authored or changed relation must still use a canonical code:
  or database:// identity.
• Stop letting R reschedule a Code authoring batch: a submitted batch is no
  longer answered with a zero-write code_candidate_relation_replan_required
  replacement plan, so a repository whose Entries reference each other across
  batches, including a mutually-referencing cluster larger than the machine
  Code batch size, now reaches aligned in the ordinary rolling batches
  instead of replanning. Receipts issued before this change still load, so a
  plan already in flight survives the upgrade.
• Keep R references from other Entries from blocking ordinary
  aoci_remove_entry orphan removal, which no longer fails with
  remove_orphan_relation_still_valid, and from making a legacy aoci.txt
  Index ineligible in aoci cognition migration snapshot when one relation
  names a path that is gone. Any dangling annotation left behind stays
  model-owned until the model's next Whole-Index read; orphan proof, Guard,
  exact preimages, and the existing Entries recovery path are unchanged.
• Report R problems only from the Entry line itself: aoci index entries check still warns about that line's own form, such as an empty item, a
  placeholder mixed with real targets, or a full-width comma, and those
  warnings still never reject an Entry; it no longer resolves targets on disk
  to warn that one is missing, duplicated, self-referencing, or not a regular
  file, and aoci cognition system relations reports no relation findings.
  The hard safety gate on the actual write path is unchanged.
• Carry the aoci scan remediation in a blocked Volumes Guide that has no
  Baseline: the response adds the scan command, a baseline_missing stop
  with cause and safe next action, and the instruction to author nothing
  before a Baseline exists. An initialized Volumes repository with zero
  Entries is completed through scan, Guide, and no-argument Maintain;
  aoci cognition bootstrap governs only an uninitialized repository or the
  exact zero-Entry Legacy minimal skeleton.
• Roll a partially written [code,database] receipt batch forward when the
  identical evidence-bound candidates are resubmitted, so an interruption
  between the Code and Database writes finishes the remaining Volume instead
  of stopping at code_candidate_plan_stale behind a pending
  .aoci/transactions receipt; the roll-forward still requires a version-4
  recovery receipt proving that Volume's own postimage, and the fixed
  Code-then-Database order, [code], and [database] batches are unchanged.
• Carry the plan-time Curation exclusions in the managed-scope-change envelope
  as curation_exclusions and replay them when verifying after publication,
  so a reviewed curation.json exclude decision for an already baselined path
  applies and archives as one transaction instead of leaving a complete
  transaction that can never be archived; envelopes without the field keep
  recomputing exclusions from the current Baseline, and envelope_version is
  unchanged.
• Keep the partition facts of mid-level tables when linking multi-level
  PostgreSQL partitioning: a table that is both a partition and a partition
  parent now records parent_object and bound while retaining
  partitioned, method, expression, and any child_objects already
  linked to it, instead of being rewritten as a non-partitioned leaf; the
  narrow non-partitioned openGauss profile and Evidence v1 are unchanged.
• Name the cause when aoci init stops with
  managed_scope_auto_authorization_blocked: the bilingual message now
  separates tracked paths excluded by a built-in safety rule (up to five
  named, the rest as a +N remainder) from a profile that assigns no path to
  the index role and from configured exact high-risk opt-ins, pointing the
  first two at aoci scope safety and --scope-profile production; --json
  reports error_code as that machine code instead of the generic config,
  and the exit code is unchanged.
• Record the effective apply authorization mode in the Managed Scope Baseline
  receipt as apply_authorization_mode, and raise
  approval_policy_relaxation when a Scope Change runs under a mode weaker
  than the one that receipt records; that risk is high, blocks
  policy_bound_auto, and forces interaction under legacy, so a team's
  review posture can no longer be lowered and self-ratified inside the same
  transaction. A receipt written before this field is not retroactively
  blocked, so the guarantee starts at the first receipt that records a mode;
  an unrecognized recorded mode fails closed.
• Accept a Managed Scope Baseline receipt recorded under the opposite
  filesystem case semantics when every managed path provably takes the same
  role and the same fingerprint participation under both, so a Baseline
  established on a case-sensitive checkout no longer stops a case-insensitive
  one with scope_change_required. The receipted value becomes the reported
  desired_policy_identity; any real case divergence leaves
  alternate_policy_identity empty and keeps the stop.
• Route every read-only Git query against a scanned repository through one
  hardened invocation that disables core.fsmonitor, core.hooksPath, and
  core.pager on the command line, so a target repository's own .git/config
  can no longer point those reads at a program for Git to run while it walks
  the working tree. Safe Inventory collection and business-source manifest
  building both use it; file listing still passes core.quotepath=false,
  every call still sets GIT_OPTIONAL_LOCKS=0, and the reported tracked,
  untracked, and ignored facts are unchanged.
• Stop the Windows-only non-atomic overwrite fallback of AtomicWrite from
  destroying the previous content by renaming the target to a same-directory
  backup before the retry, restoring that backup when the retry fails, and
  keeping both copies with both paths named in the error only when even the
  restore fails; the fallback still runs only after a normal atomic rename
  has failed and never on other platforms.
• Pin * text=auto eol=lf in .gitattributes so a source checkout is
  byte-identical on every platform: the Windows default core.autocrlf=true
  can no longer rewrite tracked text to CRLF, so a Windows checkout still
  matches the raw-byte Baseline instead of leaving aoci_maintain blocked on
  the whole tree. Every tracked text file is already LF, so the
  repository-wide renormalization changes no committed bytes.
• Report service_binary_replaced_on_disk: true in the Volumes
  aoci_maintain result, aoci_rules, and the final Overview metadata when
  the service binary's on-disk size or mtime no longer matches what the
  running process started from, so restarting the host MCP integration
  becomes a machine fact instead of manual diagnosis. The fact is advisory:
  it blocks nothing, is absent when there is no drift, counts a vanished
  binary as replaced, stays off when the startup probe records no identity,
  and does not change the nine-tool MCP surface.
• State where the running version and binary path come from in the Runtime
  Rules and the AGENTS integration block: the service version is
  cognition_receipt.mcp_service_version in any check_only Overview or
  Maintain response, the binary path is the command in the project's
  .mcp.json, and the CLI need not be on PATH, so a missing shell command
  never means AOCI is absent.
• Add three standalone black-box suites under scripts/blackbox/ that drive a
  built aoci binary as a real stdio MCP client and never import internal
  packages: mcp_conformance.py makes 58 read-only checks of the handshake,
  the nine-tool registry, input Schemas, response shapes, and malformed input;
  mcp_scenarios.py runs 30 fault-injection scenarios for cursor replay and
  tampering, write-lifecycle rejection, crash injection, and racing writers
  over disposable fixtures; and mcp_lifecycle.py takes three frozen fixture
  projects, including a 453-object one, from init through incremental
  maintenance, Database Evidence acceptance, schema drift, multi-batch
  authoring, and re-alignment, with an optional real-agent model track. Every
  suite now also asserts that no non-Overview tool response exceeds 64 KB
  under default configuration, so a response can never grow past what an
  ordinary Host displays inline. All three need only Python 3, git, and a
  binary, honor AOCI_BIN, and ship with a repository clone rather than with
  Release archives.
• Correct the public delivery contract in
  spec/public/aoci-overview-delivery-v1.txt and docs/overview-delivery.md:
  an exact replay of a genuine cursor idempotently re-serves the identical
  Chunk bytes, and because a cursor is re-derivable from its bound facts
  alone, an unchanged Index and chunk_tokens accept the same cursor across
  MCP process restarts. Invalid, missing, reordered, and cross-chain use
  still fails closed, overview_snapshot_changed and
  overview_chunk_tokens_changed still require a restart at Chunk 1, no
  delivery Session or transaction is persisted, and the Volumes governance
  binding remains in-memory session state that does not carry across
  processes.
• Correct the aoci_remove_entry description and the Volumes and system
  cognition Specs to state the behavior the machine actually has: an R
  reference from another Entry never blocks orphan removal, a dangling
  annotation left behind is model-owned semantics handled on the next
  Whole-Index read, and relation content produces no Finding, so findings
  stays in the public shape and stays empty. Removal and projection behavior
  is unchanged.
• Document the shipped meta scope of aoci_overview and the additive
  aoci_maintain input intent=cognition_optimization with its optional
  object_refs filter of canonical code: references in
  aoci-cognition-volumes-v1.txt, and cross-reference that intent from
  aoci-database-cognition-authoring-v1.txt, where it never applies to
  Database assessment. project and meta each deliver Root + Meta, no MCP
  tool is added, and ordinary no-argument Maintain is unchanged.
• Renumber the first-section items of the aoci_rules runtime-rules contract
  from 5.-8. to 4a.-4d. and prefix the en-US section headings with
  Section , so an item number no longer collides with the deliberate global
  5.-18. sequence of the later sections or with a section number. Item
  bodies stay byte-identical in both official locales.
• Stop the aoci cognition and aoci database group help from claiming
  read-only behavior neither group has: cli.short.cognition now names
  governed Apply workflows next to layout planning, since the group carries
  aoci cognition bootstrap apply and aoci cognition onboard apply, and
  cli.short.database scopes its read-only claim to database access, since
  those workflows write local Schema Evidence and advance the evidence
  Baseline. Both official locales change together; no command, flag, or
  runtime behavior changes.
• Correct the Git claim in docs/install.md and docs/supply-chain.md: the
  binary starts and non-Git directories are scanned without it, but Safe
  Inventory invokes the host git executable in any root holding .git for
  tracked and ignored authority and fails closed with
  safe_inventory_git_unavailable when that executable is absent. GitHub CLI,
  Cosign, Go, and SBOM readers remain verification-only with no runtime
  dependency.
• Document that the host configuration written by aoci init --agent <name>
  (.mcp.json, .claude/settings.json, .codex/config.toml,
  opencode.json) embeds machine-bound absolute paths and belongs in
  .gitignore, and add a moved-binary-or-repository troubleshooting entry:
  the Claude and Codex installers detect an existing entry by key presence,
  so re-running init keeps the stale paths and aoci doctor still reports
  those integrations as installed while the server fails to start, whereas
  OpenCode fails closed on an mcp.aoci conflict; recovery is to remove the
  stale mcpServers.aoci, [mcp_servers.aoci], mcp.aoci, and PreToolUse
  entries and re-run init from the new location.
• Scope the Windows Host-Agent guide to the layout it describes, marking its
  Entries/Header/Curation Stage sections Legacy, routing Volume-first
  repositories to the live Guide with ordinary no-argument aoci_maintain and
  aoci_update_entry, replacing the old aoci_overview size-threshold
  fallback with the current continuation_required chunked delivery, and
  recording the evidence-driven automatic replan, Resume, and Rollback
  closures for a stopped write. The "F length is never machine-blocked" rule
  now covers Legacy v1 Entries only, because FRAS v2 hard-limits a Volumes v1
  object's F to 160 runes.
• Add English renderings of three Chinese-only documents:
  docs/contract-authority.md, spec/public/s-field-discipline.en.txt, and
  docs/windows-host-agent.en.md, reachable from AGENTS.md,
  spec/public/README.md, docs/troubleshooting.md, and a new pointer in
  docs/windows-host-agent.md. The Chinese files remain the normative
  originals; the renderings carry no machine-scanned declaration and open no
  second rule authority.
• Expand both public READMEs with the Volume-first layout, the three-stage
  workflow, cross-Agent and cross-session reuse of one Whole-Index, and a
  verbatim excerpt of the starter tag dictionary with a worked reading of one
  compact tag; each repository's formal Meta remains authoritative.
• Build every gate and the signed release from the single go.mod toolchain
  directive, now go1.26.6, so release archives carry the same patched
  standard library that fast CI, full confidence, and the rehearsal verify.
• Preserve the nine MCP tool names and their stable identity, FRAS v2, and the
  existing Index and Baseline formats; every response change in this release
  is an additive field.

The first public availability date for v0.1.0-rc4 is 2026-08-15.

See docs/install.md and docs/supply-chain.md for asset verification.