All notable changes to this project are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

• Align devDeps pins to the published dsh 0.1.2-alpha.2 line (0.1.1-rc.2 -> 0.1.2-alpha.2); no behavior change to envelope/gating semantics.

• Tests and the loader runner no longer import the CallId brand from @deepseek-ai/dsh-llm (renamed to ToolCallId on host master): the call-id brand is now derived from the dsh-tools execution contract, staying green on both the published rc line and the 0.1.2-alpha.1 checkout.
• The adaptive-append comments now state the 0.1.2-alpha.1 reality (the ignorable envelope was removed, so the probe fails safe and the append is skipped). Behavior unchanged.

• Self-contained offline HTML report. data_report gains format: html (with key): a profile or clean report renders as a single .html file with inlined CSS/JS, no external requests (no CDN, <link>, or <script src>), the DAMA six-dimension scorecard, the per-column profile summary table, and the per-rule cleaning summary table. Dataset values are HTML-escaped before reaching the markup.
• Versioned report schema. The profile report now carries schemaVersion (a new REPORT_SCHEMA_VERSION = 1), declared in the profile report's JSON schema so future consumers can detect an incompatible persisted record instead of misreading it.

• DAMA six-dimension quality scorecard in data_profile: completeness, uniqueness, validity, consistency, timeliness, and accuracy (limited declared-schema definition; accuracy is reported undetermined without a declared schema rather than fabricated). The scorecard flows through the profile report, the persisted storage-domain record, and the tool result.
• Full-row sha256 content-hash duplicate detection in data_profile, reporting the duplicate rate plus a bounded sample of 0-based duplicate row indexes (capped by evidenceRowLimit).
• Pre-delivery contract validation summary on data_clean (dedupe before/after row comparison, uniqueness regression, non-null and type regressions) plus a persisted clean before/after profile diff report under the new clean-diff storage record kind.
• data_report tool plus ctx.dataQuality.getReport/listReports query methods to read persisted reports back by reportKey (path-safe validation, missing records fail loud) or by kind, ordered chronologically.
• data_clean dryRun parameter: no file written and no report persisted, returning the per-column cleaning plan (in contract.columnDecisions) plus the expected contract/diffPreview.
• Numeric distribution count/distinct fields on numeric column profiles.
• Config scorecardWeights (six per-dimension weights, non-negative, equal by default) driving a weightedOverall scorecard total.
• Per-column decision trace in the clean contract (columnDecisions: strategy + affected rows), persisted with the clean report.
• File-encoding reporting in data_profile (encoding: UTF-8 BOM presence and validity; invalid UTF-8 is reported, not fatal).
• Built-in industry profile presets (retail, saas, fund, real-estate, e-commerce, healthcare, logistics, manufacturing, energy) via the data_profile industryPreset parameter, feeding the scorecard accuracy dimension; unknown ids fail loud.
• Metric reconciliation on data_verify: optional expectations (rowCount/columnSum/columnMean/uniqueCount/nullCount, each with column, expected, optional relative tolerance) reconcile deterministic computed values; a mismatch is a normal passed: false verdict with actual/expected/tolerance detail, while invalid metrics, columns, and out-of-range tolerances fail loud. Results flow into the verify report and its persisted record.

Documented, deliberate non-goals for this repository (recorded rather than silently omitted):

• Native xlsx — a real Excel parser needs a new dependency (or a sidecar); this repo keeps a zero-dependency parsing contract, so CSV/TSV/JSON/JSONL are the supported dataset formats, now with UTF-8/BOM encoding detection.
• SQL / DuckDB access — database drivers are out of scope for the zero-dependency, in-process TypeScript seam; tabular datasets are files.
• qsv as an alternative parser — an external binary violates the no-external-processes contract.
• Slot client report panel — a browser report panel needs a client half and bundle changes; the consumption path is covered by the data_report tool plus storage-domain persistence instead.
• Computer-use fallback collection — macOS-only and brittle; the regular filesystem capability already covers the supported collection path.

• Persist the bounded clean preview in the durable clean report so the model-visible canonical result is fully reconstructable from its reportKey (model-visible ⟺ logged).

• Bumped the @deepseek-ai/dsh-* dev dependencies to the 0.1.1-rc.2 release line (peers stay >=0.1.0-rc.8 <0.2.0; no rc.2-only API is used) and re-verified the full gate against the published 0.1.1-rc.2 types. Compatibility metadata (dshWorkshop.compatibility.dshVersions) now advertises 0.1.1-rc.2, and the CI compat workflow pins the dsh/dsh-base/dsh-headless installs to 0.1.1-rc.2. Session-event behavior is unchanged: 0.1.1-rc.2 still exposes no plugin event-registration surface and no ignorable append option, so the adaptive gate keeps skipping the append and the storage-domain report stays the durable record.

• Bumped the @deepseek-ai/dsh-* peer and dev dependencies to the 0.1.0-rc.8 release line (peers now >=0.1.0-rc.8 <0.2.0) and re-verified the full gate against the published rc.8 types. Session-event behavior is unchanged: rc.8 still exposes no plugin event-registration surface and no ignorable append option, so the adaptive gate keeps skipping the append and the storage-domain report stays the durable record.

• The ctx.dataQuality capability seam (Service Definition / local deterministic Provider / tool Consumers) with the frozen cross-plugin verifyCitations(request) contract for document-number ↔ dataset checks (verified / mismatch / not-found / unverifiable, relative-tolerance numeric comparison).
• data_profile tool: deterministic dataset profiling (row/column counts, inferred types, missing rates, unique counts, numeric distributions, IQR outliers, mixed-type notes, duplicate-row counts) over workspace CSV/TSV/JSON/JSONL, with optional systematic sampling.
• data_clean tool: ordered declarative cleaning rules (dedupe, fill-missing, coerce-type, normalize-unit, trim, map-values) with a per-rule audit log; the source file is never overwritten and cleaned output writes only inside the workspace.
• data_verify tool: declarative verification rules (not-null, unique, range, regex, enum, cross-column, freshness) with capped failing-row evidence; overall failure is a normal passed: false result, never a tool error.
• Durable report persistence to the data_quality storage domain (keyed by run timestamp + dataset fingerprint), adaptive data-quality/profile|clean|verify session events on hosts that can carry them safely, fail-loud Schemastery config, and workspace-confined path resolution (escape + extension + size + row guards).