Arckit

2026.08.14.1

Ports three verified upstream arc-kit defects in the EU Cloud Sovereignty Framework v1.2.1 implementation into this fork, found against the Commission's own Implementation guidance PDF and Annex Sovereignty assessment calculator XLSX, plus one related backward-compatible widening. TDD RED round: the weight fix's test asserted all eight weights individually against unmodified source first, capturing SOV-1/SOV-5/SOV-7 failing by name before any production code changed — a sum-only assertion would have missed all three, since both the wrong and the right weight sets sum to 100.

• SOV_WEIGHTS had three wrong values. SOV-1 was 15 (should be 20), SOV-5 was 20 (should be 10), SOV-7 was 10 (should be 15) — ground truth is guidance p.7 / calculator cells D4/D45/D76/D102/D133/D169/D195/D231. The wrong set summed to 100 too, which is exactly why it survived review. Previously-written sovereigntyAssessment records were scored against the wrong weights and are NOT recomputed by this upgrade — that resource is lifetime: "infinite", so such records persist; re-run euSovereigntyScore for any assessment that still matters.
• SEAL_LABELS.SEAL3.en was wrong. Corrected "Digital resilience" → "Technological sovereignty" (guidance p.2-3, p.10). SEAL_LABELS.SEAL3.nl is UNCHANGED — "Digitale veerkracht" is a verified, deliberate divergence from the Commission's English name, quoted verbatim from the NDS Cloudprogramma notitie "Verkenning Overheidsbrede Soevereine Clouddiensten" (11 juni 2026), Tabel 1 p.8 — not a bug, and not something a future pass should "fix" to match the English.
• No overall SEAL was computed, even though it is the framework's actual rejection gate (guidance p.9: "The overall SEAL level is the lowest SEAL level achieved in any of the objectives" — calculator cell F2: ="SEAL-"&MIN(H5:H251)). computeSovereigntyScore now returns overallSeal (a MINIMUM across all eight objectives, never an average or a mode; undefined — never fabricated as SEAL0 — when any objective has no recorded SEAL) and overallSealGovernedBy (every objective id tied at that minimum). SovereigntyAssessmentSchema gains both fields.
• maxScore had no default and no documented ceiling. Added an exported SOV_MAX_SCORES constant with the calculator's actual per-objective maxima (SOV-1 1000.03 · SOV-2 1002 · SOV-3 1000 · SOV-4 1002 · SOV-5 1001 · SOV-6 1000 · SOV-7 1001 · SOV-8 1000 — a nominal 1000 is design intent, not arithmetic; the workbook rounds each criterion's answer value to 2dp). objectives[].maxScore is now OPTIONAL: when omitted, the accept/reject GUARD widens to the objective's actual SOV_MAX_SCORES[id] ceiling (a flat-1000 clamp would have falsely rejected a legitimate maximal SOV-2 response of 1002), while the contribution DIVISOR stays the calculator's flat nominal 1000 — the same formula the calculator itself uses for every objective regardless of that objective's true ceiling. Consequence, kept faithful rather than hidden: a maximal response across all eight objectives (every objective at its SOV_MAX_SCORES ceiling, maxScore omitted throughout) scores 100.0756%, not 100%. An explicit caller-supplied maxScore is honoured exactly as before for both roles — the new defaulting only applies when it is omitted.

manifest.yaml and model.version bumped in lockstep with a fifth identity upgrades[] entry — additive/corrective resource-shape changes only, no data transformation. New suite file arckit_workspace_csf_test.ts (26 tests) plus corrections to three stale assertions in arckit_workspace_nl_test.ts that hard-coded the old wrong weights (a local SOV_WEIGHTS duplicate, one sanity-test title, one hard-coded 15 in the fractional-contribution test). Suite is now 274, all green; deno lint/deno fmt --check clean.

2026.08.07.1

Release of the NL sovereign-cloud overlay, with two defects closed that the pre-publish adversarial review found in 2026.08.06.1. 2026.08.06.1 was never published — the registry moves 2026.08.02.1 → 2026.08.07.1 directly. See the 2026.08.06.1 section below for the overlay itself.

• computeSovereigntyScore's fail-closed guard was one-sided. It rejected a missing objective but silently accepted a duplicate id — a repeated SOV-1 changed the computed score from 50 to 57.5, the later entry winning via the new Map(...) construction — and silently ignored an unrecognized id. The objective set is now validated closed and exact (SOV-1..SOV-8, each exactly once) before scoring, naming the offending id.
• MigrationSchema.ladder was required with no default, so a classificationMigration record written before the field existed failed to parse. That resource is lifetime: "infinite", so such records persist and can be restored from the datastore. It now defaults to "uae" — the only ladder that ever existed — matching the .default([]) precedent set for skipped/unmappedFiles. MigrationSchema is exported so a test pins it.

Both passed a fully green 244-test suite: the tests only exercised the missing-objective direction and never parsed a legacy record. Suite is now 248. quality.yaml gains its entry and lists arckit_workspace_nl_test.ts under the coverage role.

2026.08.02.1

Real-fix pass closing all six remaining latent bugs tracked on the LOCAL arckit-latent-bugs issue-lifecycle model (NEVER filed to the swamp.club Lab) — LB2 (MEDIUM) and LB3..LB7 (LOW). arckit_workspace.ts was no longer byte-frozen after the LB1 fix in 2026.08.01.1; this is the second production change.

• LB2 (MEDIUM) — migrateClassification apply=true non-atomic overwrite, no backup. The apply branch now copies the pre-migration content to a <artifact>.bak recovery sibling, writes the proposed text to a sibling temp file, then Deno.renames it over the artifact — an atomic replace on the same filesystem, so a reader always sees the whole old or whole new file, never a truncated partial, and a crash mid-write leaves the artifact intact plus a recoverable .tmp orphan.
• LB3 (LOW) — unbounded readTextFile in template / migrateClassification. Added a defaulted global arg maxFileBytes (default 10 MiB) to GlobalArgsSchema. migrateClassification cap-checks via the scan snapshot's sizeBytes WITHOUT reading, recording an oversize artifact in the new skipped[] field (reason: "oversize") in both report and apply modes; template Deno.stats the bundled file and rejects before reading if it exceeds the cap.
• LB4 (LOW) — non-enum projectState.state vacuously satisfies the gate. ProjectStateSchema.state is now z.enum(PROJECT_STATES) (PHASES plus complete/abandoned) instead of z.string() — a schema seam change only. readProjectState (the sole reader for status/advance/skipPhase/abandon) now rethrows a friendly Corrupted project state for <dir>: ... error instead of a raw ZodError when a hand-edited/datastore-restored state falls outside the enum, so advance can no longer auto-complete a bogus phase. nextPhase and gateFor/evaluateGate are UNCHANGED pure functions — the frozen contract-fixture (nextPhase("bogus") === "complete") and coverage_test.ts's gateFor/evaluateGate bogus-phase assertions stay byte-behaviorally identical; the schema simply makes them unreachable from the public API. Trade-off: abandon on a corrupted record now also rejects (fail-closed, cannot rescue a corrupted record via abandon) — accepted for a LOW corruption-recovery path.
• LB5 (LOW) — project-id allocation breaks past 999. parseProjectDir's regex and the startProject allowlist guard both widen \d{3} to \d{3,} (3-OR-MORE digits); nextProjectDir's padStart(3, "0") is unchanged, so ids <=999 keep their existing 3-digit zero-padding and ids >=1000 widen naturally. The guard's character class still forbids /, \, ., so every LB1 traversal payload stays rejected; ARTIFACT_RE (artifact filename ids) is untouched and stays 3-digit.
• LB6 (LOW/info) — templates vs provisionTemplates inventory divergence. templates now additionally walks the bundled templates/ directory (the same source provisionTemplates copies) and surfaces any file with no TEMPLATE_MAP command in a new defaulted unmappedFiles: string[] field on TemplateCatalogSchema, reconciling the two methods' inventories. templateCount and the sizeBytes:0 missing-file behavior are unchanged.
• LB7 (LOW/info) — symlinked artifacts silently skipped. listFilesRecursive and scanWorkspace now resolve a symlink entry's target kind via Deno.stat — a symlinked artifact or project directory is inventoried like a real one (bounded by the existing depth cap). Write-safety cross-cut with LB1/LB2: migrateClassification's apply branch Deno.lstats before writing and skips (reason "symlink", reusing the LB3 skipped[] field) rather than writing through a symlink to a target outside the workspace; report-only mode still reads through the symlink and proposes.

extensions/models/arckit_workspace.ts: all six fixes land in a single consolidated ordered block inside migrateClassification's per-file apply loop (LB3 cap-check → read+propose → LB7 symlink skip / LB2 backup+atomic write), plus the schema/regex/directory-walk changes above. manifest.yaml/model.version bumped 2026.08.01.1 → 2026.08.02.1 in lockstep. Added upgrades[] (previously absent on this model) with identity upgradeAttributes: (old) => old entries — maxFileBytes is a defaulted global arg and skipped/unmappedFiles are defaulted resource-schema arrays, so no stored data needs transformation.

• extensions/models/arckit_workspace_adversarial_test.ts: flipped LB2 (:121), LB4 (:197, :221), LB5 (:251), LB6 (:270), and LB7 (:306) from pin (arckit-latent-bugs LBN, SEVERITY):-titled current-behavior pins to fix regression (arckit-latent-bugs LBN, SEVERITY):-titled POST-fix assertions; relabeled the two LB3 500 KB pins from "

Release 2026.07.16.2 — align model versions with manifests

Maintenance release across the @magistr extensions. For most packages this carries no functional change: the only edit is the model's version: field, brought back in line with its manifest version so the published model type version and the package version no longer drift.

Functional changes in this release are limited to:

• anime-cron: normalizeTitle now strips a ": subtitle" suffix and a trailing parenthesized year before comparison, fixing dedup false-misses where the torrent title carries a subtitle or year that the AniList romaji does not.

• arckit: first publish. Standalone ArcKit port — a 12-phase architecture governance state machine with 65 bundled templates, driven by a bundled skill.

Also tracks three extensions (kaiten, observability-agent, music-library) that previously existed only as untracked working-tree directories, recovered from stashes.