v0.2.0 — reliability prerelease

Marginalia v0.2.0, reliability prerelease.

Builds on the v0.1.0 agent-facing MCP surface. No MCP tool or parameter changes.

Providers

• marginalia provider login chatgpt: interactive device-code login into Marginalia's own credential directory. Refuses ~/.codex and ~/.pi/agent (symlinks too), a non-TTY session, and an existing credential unless --force is given.
• marginalia provider status chatgpt [--json]: read-only report of plan, account, subscription and token expiry; exits 1 when the credential is unusable.
• Transient provider errors are now retried in ask synthesis and in every ingest judge and curator step (curator, relation curator, merge judge, predicate resolution, type adjudication, correction judge, predicate-propose sweep judge). Two attempts, retryable errors only, Retry-After capped at 60 s, each retry recorded. A single 503 used to turn an answer into empty text or silently degrade the step. The reconcile cluster judge stays single-shot on purpose.

Vaults that stopped accepting writes

• A case-variant re-mention of an existing entity no longer wedges the vault. The sealed-plan applier compared raw titles while the node id folds case, so one capitalisation variant made every later remember fail.
• Committing a parked manual-review item no longer wedges when another document already committed the same node. A vault the old code wedged heals by re-running the same review resolution.
• Every remember write failure is logged on both MCP and REST: graph-write failures at ERROR with cause, caller mistakes at WARNING. REST previously logged nothing. Response codes unchanged.

Other

• Neo4j backend: DriverError is translated to GraphBackendError at every driver entry point (reads, writes, wipe, Neo4jStaging).
• The bundled web UI was rebuilt to match its source.
• Private network addresses removed from published docs, including the README that ships as the wheel's long description.
• LoCoMo harness (source repo only): --enable-chatgpt, --ingest-concurrency for parallel conversation ingest, and ingest_units_failed in summaries, reports, compare and the ledger.

Verification, stated plainly

Source GitHub Actions did not run on the source commit. All four workflows were triggered and every job failed without starting: "The job was not started because recent account payments have failed or your spending limit needs to be increased." The gates were reproduced locally on that exact commit instead:

• tests: uv lock --check, actionlint, ruff clean; canonical CI pytest selection 1 failed / 4440 passed / 65 skipped / 37 deselected / 1 xfailed (the failure is the known issue below); npm ci, audit (0 vulnerabilities), typecheck and build clean; no frontend_dist drift; docs build leaves the tree clean.
• acceptance: every scenario PASS on grafx (54 skipped by design on that backend).
• eval floor: gold 32/32, distractors 20/20, provenance gate pass, recall_floor selftest PASS, regressions [].
• neo4j backend gate against a live neo4j:5-community: contract tests 9 passed / 0 skipped, CI pytest step 1 failed (same known issue) / 1394 passed, scenario 84 PASS with 25 assertions.
• Against this exact wheel: dependency contract 17 passed, install footprint 1 passed, live [serve] start from an isolated HOME with zero vaults answering /health, /version and the SPA, JSON-LD export through [jsonld], every extra in its own clean Python 3.12 env, both CLIs, retired constructors failing closed.

Not performed for this release:

• The Windows managed-credentials (DPAPI) gate needs a Windows runner and did not run.
• The interactive Windows PowerShell 5.1 lifecycle rehearsal was not performed. No published version has passed it.

Known issue, pre-existing, not fixed here: tests/store/contract/test_snapshot_concurrency.py::TestLadybugSnapshotPinnedDumpVsSwap::test_dump_never_mixes_pre_and_post_swap_rows fails under load. It is a race between the Ladybug staging commit's two renames and LadybugStore.snapshot() opening graph.lbug read-only. A deterministic probe reproduces it identically on v0.1.0. Waived for this prerelease.

Source: 314e3af15a901c71c0bcb7af75a63f081298859f
SHA-256: 59adee4483d8f85525e30a5e6271e6328238ac275d4b3fe05e29325d2ed29520

v0.1.0 — agent-facing MCP surface

Marginalia v0.1.0 — agent-facing MCP surface.

First release with a real public tool surface: five MCP tools and seventeen ask parameters, up from four tools and three.

Vault selection

• New list_vaults tool — names only, no filesystem paths and no path-derived ids.
• Optional per-call vault=<name> on ask, explore and remember. Registry names only; a path-shaped value is refused. The connection's ?vault= keeps precedence, and a discarded override is reported as vault_override_ignored.
• ask and explore now report which vault served the response.

Retrieval control

• ask exposes the retrieval policy the web UI already had: enable_subgraph, source_block_policy, seed_k, max_degree_per_seed, neighbour_budget_tokens, source_block_budget_tokens, coverage_threshold, min_claim_confidence, max_nodes, max_relationships, max_claims, relationship_types, plus include_sources.
• enable_subgraph is opt-in and its default is unchanged: a grounded eval measured block-dump answers at 0.792 against the subgraph path at 0.6.
• explore gains relationship_types, min_claim_confidence, max_degree_per_seed, emits block_id on claims and relationships, and returns its own retrieval block.
• include_sources returns block_id, byte range, content_hash and a vault-relative path.

A degraded answer no longer looks like a good one

• synthesis_status is always present: ok | empty | truncated | abnormal_stop | provider_error. An empty text with provider_error means the model was unreachable — not that the graph lacks the answer.
• finish_reason and native_finish_reason are surfaced; anything other than stop is treated as suspect.
• Reasoning that the provider inlines into the answer is now stripped even when the opening tag never arrives.
• Assembled context marks each excerpt with its path, byte range and file size, so the model can say the excerpts do not cover a period instead of inferring a value.

Hardening

• No filesystem path reaches a client-facing error: a vault-name length cap enforced before any filesystem access, one sanitiser for OSError at the resolution and lease seams, ELOOP handled through the RuntimeError pathlib raises for it, and no resolved-path echo on the absolute-path branch.
• The MCP handshake reports Marginalia's version instead of the framework's.
• A 401 now distinguishes a missing header, an empty credential, an unsupported scheme and an invalid token. The unauthorized error code is unchanged.

Compatibility

Additive. New tool, new optional parameters that default to inheriting existing behaviour, new response fields. The only behaviour change is the 401 detail text; anything matching on the unauthorized error code is unaffected.

Verification, stated plainly

Passed, reproduced locally: 4052 tests passed / 94 skipped / 1 xfailed, ruff clean, uv lock --check clean, docs gate 47 passed. Against this exact wheel: dependency contract 17 passed, install footprint 1 passed.

Not performed for this release:

• GitHub Actions did not run on the source commit — the source organisation's Actions are billing-blocked. Every source gate above was reproduced locally instead.
• The Neo4j backend gate runs only in source CI, so it did not run. The Neo4j surface is unverified for this release.
• The interactive Windows PowerShell 5.1 lifecycle rehearsal was not performed.

Known, pre-existing, not fixed here: declining the installer's first-run vault prompt aborts the install and rolls it back. Pass --no-onboard to skip the prompt.

Source: 8e0e9ae2c51e6c452705548dcba5af8fd77d909c
SHA-256: 185d787947d524f6fd7a88d120c6bbbff1d36d55348eb18d11bdf2c65bce8591

v0.0.50

Prerelease: Linux Docker+tmux rehearsal pending; Windows lifecycle rehearsal deferred.

Wheel marginalia-0.0.50-py3-none-any.whl
SHA-256 cc524877686ad0227d6cd208c487f027acf1f1e25772c98afe00b96bbdf3ea1e
Source 5264b971dc3e97390a25cff06b8ff0d1dcfeb04d

Marginalia 0.0.49 (prerelease)

Public distribution for source tag v0.0.49
(85f2b28bd1c23c1cdc5fdf6389d7ba96862bda6b).

Changes since 0.0.48

• The onboarding base URL is now canonicalized in one resolver, regardless of the form typed.
  Discovery used to probe {base}/models (treating user input as the server root) while completions
  POSTed {base}/chat/completions (treating input as already versioned), so a server root like
  http://host:port passed discovery but 404'd on the completion step, while a /v1-form base
  failed discovery on servers that only serve /v1 (oMLX). Every input form (root, root/, /v1,
  /v1/, subpath variants) now derives the identical pair — discovery at {root}/v1/models,
  completions at {root}/v1 — and the base persisted to marginalia.yaml is always the canonical
  {root}/v1, so the saved config never depends on which form was entered. A query string on the
  base (Azure-style ?api-version=) is preserved on every derived URL; non-OpenAI-contract drivers
  (Anthropic, Gemini, Azure, Ollama, …) are untouched.

• Onboarding runs a real completion before it saves anything. After discovery lists models, a
  minimal completion runs through the exact canonical endpoint every real ask/ingest call uses. On
  failure the run aborts with the exact attempted URL and nothing is saved — no LLM config block, no
  env secret — instead of persisting a base the runtime could not actually use.

• A non-interactive run without --model never silently defaults to the first discovered model.
  On a multi-model server models[0] can be a non-chat model; the run now lists what it found and
  demands an explicit --model (a preset's own declared default is still honored when it is among
  the discovered models).

  Migration consequence. Existing vaults keep their pinned config; for OpenAI-compatible drivers
  the stored api_base is now canonicalized to the runtime's own shape, so a previously saved
  root-form base now resolves to the same /v1 endpoint a fresh install would get — previously
  broken root-form configs start working, and correctly saved /v1 configs are unchanged.

Verification

GitHub Actions is unavailable for the private source repository (billing not enabled), so the
marginalia CI jobs did not run for this release. Every required gate was verified locally against
the exact release inputs:

• model-free-tests / tests -> green on release-candidate commit 2649f2a: 3,762 passed, per that
  commit's recorded verification.
• eval-gate / floor -> run locally exactly as the workflow invokes it, against source SHA
  85f2b28bd1c23c1cdc5fdf6389d7ba96862bda6b (manifest.code.git_rev / dirty: false confirm it).
  Provenance gate: 32/32 gold targets, 20/20 distractors resolved, pass. Recall-floor selftest:
  pass. Recall-floor gate against the committed baseline: hard-recall@10 0.4688 (15/32),
  extraction-completeness 0.5625 (18/32), MRR@10 0.4348, precision@10 0.2 — regressions: [].
• release-artifact-gate / wheel -> clean wheel build from the exact tagged SHA; wheel
  metadata/UI inspection passed; dependency-contract (17/17) and footprint (1/1) against this exact
  wheel; base install plus every advertised runtime extra (embeddings, ladybug, jsonld, mcp, serve,
  litellm, bedrock, sentence-transformers) resolved and smoke-tested in its own clean Python 3.12
  environment.
• release-artifact-gate / windows-managed-credentials -> could not be reproduced. This job
  proves a DPAPI CurrentUser credential round-trip; DPAPI is Windows-only and this release was
  validated from macOS. Not covered by any local substitute in this release.
• Public marginalia-dist distribution-gate -> runs on push of the dist bake commit (pending on
  publication).
• Linux Docker+tmux public raw-URL rehearsal for 0.0.49 (including the new verify-before-save
  onboarding) -> runs after publication against the exact dist-main commit; will be recorded in the
  dated section of the marginalia-dist README with retained transcripts under
  evidence/v0.0.49/ in that repository.

This is a prerelease: no published Marginalia version, including 0.0.45 through 0.0.48, has yet
passed a real interactive Windows PowerShell 5.1 release-lifecycle rehearsal.

Artifact

• Wheel: marginalia-0.0.49-py3-none-any.whl
• SHA-256: 8274abea746e9ec6d1b8450e5d416ea626ec8e325effbe1f240929ad9ec9d4c3
• Built from the clean tagged SHA above; dependency-contract (17/17) and footprint (1/1) re-run
  against this exact file and passed.

Marginalia 0.0.48 (prerelease)

Public distribution for source tag v0.0.48
(1242306425e4245bd4292ad24d6280e310ef8d26).

Changes since 0.0.47

• The stale default LLM model is removed — the model default is now empty and discovery-first.
  The hardcoded Qwen3.6-35B-A3B-oQ4-fp16-mtp default (which assumed a local Qwen server
  that most machines do not run) is gone. The config defaults (config/_vault.py), the
  marginalia onboard legacy-local preset, the web UI (rebuilt frontend_dist), and
  GET /api/v1/config/defaults now all ship an empty model default, so a fresh install
  discovers what is actually reachable instead of assuming one. The model-free acceptance
  scenarios 83/84/98 now pin llm.enabled=false, keeping them deterministic.

  Migration consequence. Fresh installs get discovery-first model behavior. Existing
  vaults keep their pinned config: a model already recorded in a vault's config is untouched
  by this change, so there is no behavior change for any configured vault.

• The public installer now offers a conditional first-run onboarding on greenfield installs.
  On a fresh interactive terminal with no existing vault or config (greenfield + TTY), the
  installer asks once, after installing the tool and before starting the app: Y (or Enter)
  runs the terminal marginalia onboard flow (you name your own vault in-flow; the installer
  never creates one), n or EOF keeps the application-first path. The prompt is skipped
  silently for piped/CI installs (no TTY), MARGINALIA_NO_OPEN=1, MARGINALIA_VAULT
  preseeding, the new --no-onboard flag, and every reinstall or upgrade — it only appears
  while the Marginalia home is greenfield, so upgrades never prompt. Six docker-tmux harness
  scenarios cover the matrix; 18/18 pty checks pass.

Verification

GitHub Actions is unavailable for the private source repository (billing not enabled), so the
marginalia CI jobs did not run for this release. Every required gate was verified against the
exact release candidate:

• model-free-tests / tests -> green on release-candidate commit e5b2316: 3,751 passed.
• model-free-tests / acceptance -> green on e5b2316: 33/33, including the Neo4j scenario.
• docs-gate / gate -> green on e5b2316: 47 generated pages, no unstaged drift.
• ruff check + uv lock --check -> clean on e5b2316 and on the bump commit.
• eval-gate / floor -> judge.py floor, recall_floor.py selftest and gate against the
  committed frozen synthetic-ci vault, and judge.py manifest — run locally exactly as the
  workflow invokes them, against source SHA 1242306425e4245bd4292ad24d6280e310ef8d26
  (manifest.code.git_rev / dirty: false confirm it). Provenance gate: 32/32 gold targets,
  20/20 distractors resolved, pass. Recall-floor selftest: pass (byte-stable,
  regression-sensitive). Recall-floor gate against the baseline: hard-recall@10 0.4688 (15/32),
  extraction-completeness 0.5625 (18/32), MRR@10 0.4348, precision@10 0.2 — regressions: []
  against the committed baseline. This frozen fixture is opened read-only and not re-ingested.
• release-artifact-gate / wheel -> clean wheel build from the exact tagged SHA; wheel
  metadata/UI inspection passed; dependency-contract (17/17) and footprint (1/1) against this
  exact wheel; base install plus every advertised runtime extra (embeddings, ladybug, jsonld,
  mcp, serve, litellm, bedrock, sentence-transformers) resolved and smoke-tested in its own
  clean Python 3.12 environment.
• release-artifact-gate / neo4j-backend-gate -> green on e5b2316 (acceptance scenario 84
  against a disposable neo4j:5-community container).
• release-artifact-gate / windows-managed-credentials -> could not be reproduced. This
  job proves a DPAPI CurrentUser credential round-trip; DPAPI is Windows-only and this release
  candidate was validated from macOS. Not covered by any local substitute in this release.
• Public marginalia-dist distribution-gate -> runs on push of the dist bake commit
  (pending on publication).
• Linux Docker+tmux public raw-URL rehearsal for 0.0.48 -> green on 2026-09-13
  against the exact dist-main commit 6de65ad19df4dfd4d3d813e707010f6f6b3c5d84 (the
  pushed tip of main): the exact-commit release-lifecycle run (thirteen
  RELEASE_LIFECYCLE_*_OK markers plus the final DOCKER_TMUX_RELEASE_LIFECYCLE_OK,
  each exactly once, tmux pane status 0) and the new one-shot greenfield first-run
  prompt run (prompt -> Y -> marginalia onboard -> user-named vault, provider
  skipped -> daemon up with UI :7777 + MCP :8201 at version 0.0.48 -> clean
  stop). Recorded in the dated v0.0.48 Linux release rehearsal (exact driver commit)
  section of the marginalia-dist README, with the retained transcripts under
  evidence/v0.0.48/ in that repository.

This is a prerelease: no published Marginalia version has yet passed a real interactive Windows
PowerShell 5.1 release-lifecycle rehearsal.

Artifact

• Wheel: marginalia-0.0.48-py3-none-any.whl
• SHA-256: 884a6721590583eda836dcb127b0e0bb729557db461b27100a3a8e6c99808b24
• Built from the clean tagged SHA above; dependency-contract (17/17) and footprint (1/1)
  re-run against this exact file and passed.

v0.0.47

Public distribution for source tag v0.0.47
(ec523eced42e6616de825ee5d8df9e6c61dfaf0c).

Changes since 0.0.46

• Default ingest chunk size halved: 12,000 bytes to 6,000 bytes.
  ingest.chunk_size_bytes (IngestConfig and the ingest parser's
  DEFAULT_CHUNK_SIZE_BYTES) now defaults to 6,000 instead of 12,000. This is
  evidence-based: three repeats per configuration, manually adjudicated
  against a 103-fact anchored corpus, measured mean fact-recall 0.4045
  (sd 0.0148) at 6000/0 versus 0.2362 (sd 0.0056) at 12000/0 — a 0.1683 gap,
  11 to 30 times either configuration's own standard deviation, with no
  overlap across any of the six runs.

  Migration consequence. Vault config resolves ingest.chunk_size_bytes
  fresh from this default on every load — nothing is frozen into a vault's
  marginalia.yaml at creation time. So any vault that has never explicitly
  pinned ingest.chunk_size_bytes will, on the next add/edit/folder-watch
  touch of an already-ingested document, find its stored Blocks' chunking
  facet (12000) no longer matches the live config (6000). That orphans the
  existing Blocks for that document, and the whole document is re-chunked
  and re-extracted from scratch on that touch — not just the changed hunk.
  Existing stored Blocks and their byte anchors are untouched until then;
  this is not a retroactive rewrite, and kg rebuild always uses the live
  config regardless of what any vault has pinned.

  To keep the previous chunking behavior, pin the old value explicitly
  before your next touch of an ingested document:

  curl -X PATCH http://127.0.0.1:<port>/api/v1/config \
    -H 'content-type: application/json' \
    -d '{"ingest": {"chunk_size_bytes": 12000}}'

  or write it directly into the vault's marginalia.yaml:

  ingest:
    chunk_size_bytes: 12000

Verification

GitHub Actions is unavailable for this org (billing not enabled). CI did not
run for this release. Every required CI gate was instead reproduced locally
against the exact release candidate:

• docs-gate / gate -> uv run python docs/build_knowledge_base.py, staged
  with no unstaged drift.
• eval-gate / floor -> tests/golden/bin/judge.py floor,
  tests/golden/bin/recall_floor.py selftest and gate against the
  committed frozen synthetic-ci vault, tests/golden/bin/judge.py manifest
  — run locally exactly as the workflow invokes them, against source SHA
  ec523eced42e6616de825ee5d8df9e6c61dfaf0c (manifest.code.git_rev /
  dirty: false confirm it). Provenance gate: 32/32 gold targets, 20/20
  distractors resolved, pass. Recall-floor selftest: pass (byte-stable,
  regression-sensitive). Recall-floor gate against the baseline: hard-recall@10
  0.4688 (15/32), extraction-completeness 0.5625 (18/32), MRR@10 0.4348,
  precision@10 0.2 — regressions: [] against the committed baseline. This
  frozen fixture is opened read-only and not re-ingested, so it is
  unaffected by this release's chunk-size default change.
• model-free-tests / tests -> uv run --extra ladybug --extra litellm python -m pytest -m 'not slow and not realmodel and not perf and not footprint and not acceptance_private_corpus and not acceptance_judge' -q.
• model-free-tests / acceptance -> ./bin/acceptance.sh, unfiltered.
• release-artifact-gate / wheel -> clean wheel build, dependency-contract
  and footprint tests, base install plus every advertised runtime extra
  resolved and smoke-tested in its own clean Python 3.12 environment.
• release-artifact-gate / neo4j-backend-gate -> a disposable
  neo4j:5-community container plus the [neo4j]-extra contract suite and
  acceptance scenario 84, run against it locally.
• release-artifact-gate / windows-managed-credentials -> could not be
  reproduced. This job proves a DPAPI CurrentUser credential round-trip;
  DPAPI is Windows-only and this release candidate was validated from
  macOS/Linux. Not covered by any local substitute in this release.

This is a prerelease: no published Marginalia version has yet passed a real
interactive Windows PowerShell 5.1 release-lifecycle rehearsal.

Artifact

• Wheel: marginalia-0.0.47-py3-none-any.whl
• SHA-256: b9c62fb2e8690c3f62b0ad26b3bf84535fd245e8dc2c2a2171bb7ef2b2dd8d14
• Built from the clean tagged SHA above (not the Phase 2 validation build); dependency-contract
  (17/17) and footprint (1/1) tests re-run against this exact file and passed.

🤖 Generated with Claude Code

https://claude.ai/code/session_018vER46yy59HeeU2vLDgXSk

v0.0.46

Public distribution for source tag v0.0.46
(f4e3ce4e337374257e1a007ece2ab25524633aa2).

Changes since 0.0.45

• Backend selection parity for REST and MCP vault creation. REST's
  api_vault_create and the MCP init_vault tool now default new vaults
  to the same backend as the CLI (Okto Grafx, D-94), and reject a
  non-loopback storage_uri unless allow_remote_db is set, mirroring
  the CLI's --allow-remote-db consent gate.
• Graph backend visibility. Each vault's backend now shows up in
  marginalia status, marginalia vault list, GET /api/v1/vaults,
  GET /api/v1/status, and the web UI's vault manager.
• Onboarding backend prompt. Interactive marginalia onboard now
  asks which graph backend to use (Grafx, Ladybug, or Neo4j) when
  --backend wasn't passed explicitly; non-interactive onboarding is
  unchanged.
• Grafx/Neo4j reembed fix. kg reembed no longer raises
  EmbeddingDimMismatch on its own live-graph read-back when a vault's
  embedder dimension changes.

Artifact

• Wheel: marginalia-0.0.46-py3-none-any.whl
• SHA-256: b5c3a825f7b734db62b2774c8a8cfb1b79ebfa23b8580d05619ae0b23dc5ab81

This is a prerelease: no published Marginalia version has yet passed a
real interactive Windows PowerShell 5.1 release-lifecycle rehearsal.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FfmrJL7VYn8zhyb27Q5pnH

v0.0.45

Pluggable graph backends (ADR 0041): the storage-and-retrieval seam is now a
GraphStore/IndexStore connector pair with three gating backends.

• Okto Grafx is now the default backend, installed out of the box. A
  post-M6 owner decision (D-94) promoted it from the earlier experimental,
  --accept-experimental-gated path to Marginalia's non-experimental
  default graph backend; no opt-in flag is required.
• Ladybug remains fully supported and selectable (--backend ladybug).
• Neo4j remains selectable as a server-backed connector (--backend neo4j,
  [neo4j] extra).
• LoCoMo quality parity holds across all three backends within judge noise
  (ADR 0041, M6 Parity).

Wheel: marginalia-0.0.45-py3-none-any.whl
SHA-256: 8a10b4d65d04e70a4612aba6788547c2a51bd31f3e9de446494d271b5cb7aaa9

This release stays a prerelease. No published Marginalia version has ever
passed a real interactive Windows PowerShell 5.1 release lifecycle, and
0.0.45 inherits that gap unchanged; the Linux Docker+tmux rehearsal has not
yet been run against this exact bake either.

🤖 Generated with Claude Code

https://claude.ai/code/session_0197N5yJGQadKWYXMgePyLiS

v0.0.44

Marginalia 0.0.44 (prerelease).

Source tag v0.0.44 = 07c2fd1e76a3ed562bbad46cb23028c27bc34295 (private OktoLabsAI/marginalia).
Wheel SHA-256: fbab50524b107436c19b0f790d10358789eae45222d0bf3ebe4ba5b79af8fed1 (1,022,801 bytes, 157 members).

Contents: ADR 0039 ingest correctness (integrity fence, generation sidecar, lease-guarded heal), ADR 0040 semantic graph quality, the effective-LLM-concurrency capacity owner, the three-tier laptop quality gate, judge rubric v2, and the reinstated search_claims answer partition.

This stays a prerelease. No published Marginalia version has ever passed a real interactive Windows PowerShell 5.1 release lifecycle; 0.0.44 inherits that gap unchanged. Promotion waits on that evidence.

Marginalia v0.0.43

Stable Marginalia 0.0.43 release.\n\nSource: e15d65a7c3baf018ade42910e64947dad882e9d5\nWheel SHA-256: 2ca924eadbad3819a32679fb4f2076e251f12921080c7510d281147ffad1ee44\n\nIncludes app-first browser launch, application-scoped multi-vault management and deletion, keyless private-LAN model configuration, clean-wheel dependency closure, and slower-host startup readiness.