LineLoom v1.4.2 — Local Code Quality and FIM Preview

LineLoom v1.4.2 — Local Code Quality and FIM Preview

GitHub-only pre-release (not a stable release). This preview is a feedback
candidate beyond the separately published v1.3.0 stable extension and v1.4.1
GitHub preview. It retains the existing SanskarCodes.lineloom extension ID
and MIT license. Neither Visual Studio Marketplace, Open VSX, GitHub Packages,
nor npmjs.com is updated by the GitHub publishing workflow.

New since v1.4.1

• HTML, CSS, Lua opt-in autocomplete — #144:
  three experimental language modes; the original nine default-enabled
  languages remain unchanged. Lua's editor mode requires an appropriate
  installed language provider. Models are still installed and selected by
  the user; no automatic model download or new remote endpoint approval.
• Native local LineLoom Code FIM request routing — #145:
  explicitly selected lineloom-code or lineloom-code:<tag> sends bounded
  prefix/suffix requests to Ollama when no user-selected related files are
  included. This is a routing convention not a trained or bundled model.
  Unknown models retain guarded instruction prompts.
• Offline local LoRA adapter integrity receipts — #143,
  hardening #147:
  bounded SHA-256 digests, fixed training metadata and checked file descriptors
  for developer-approved local experiments. Checksums are not provenance,
  cryptographic signatures, or proof of dataset rights.
• Real extension-host regression coverage — #148,
  #149:
  mocked Ollama requests verify opt-in language gates, model overrides,
  the native FIM prefix/suffix contract and generic fallback. Lua's
  physical-editor test is conditional on installed language mode support.
• More trustworthy synthetic LineLoom Code research —
  #150,
  #151:
  benchmark --mode auto now matches the extension's special local FIM tag,
  and reports use incompatible-by-design lineloom-synthetic-v4 protocol.
  The offline comparator matches scope/request digests against the
  checked-in public synthetic fixtures, rejecting invented hashes
  and inconsistent language/category attempt allocations.
• Source-free regression review — #152:
  allowlisted heuristic flags for additional request failures, fewer strict
  matches (including per-language), completion-format artifacts and meaningful
  completed-only p95 latency increases. This is a research aid, not
  code correctness, model quality, or evidence of an authentic run.

Verification and limitations

The GitHub-only owner-triggered publisher must pass a locked Node build, all
TypeScript/lint/unit/model/data/QA/site/source/security checks, a mocked Linux
VS Code extension-host test, complete-toolchain high/critical and production-only
moderate dependency audits, a genuine preview-channel VSIX with SHA-256,
a clean isolated VS Code CLI install and independently pinned upgrades from
the published v1.3.0 stable and v1.4.1 preview. Publishing is rejected if
the v1.4.2 tag or release already exists.

Manual QA outstanding — NOT RUN: real localhost Ollama suggestion quality,
actual GPU/CPU model latency, independently approved training data and base
checkpoint, training first-party LineLoom Code weights, physical Windows/macOS/
Linux keyboard/screen-reader acceptance, and real remote consent/privacy
testing. Simulated tests cannot substitute for these checks.
No first-party model weights, trained adapter, benchmark conclusions, dataset,
or full-repository index are bundled or published.

Development dependency advisories: existing moderate development-tool
transitive advisories remain subject to security review #79.
A passing high-severity audit does not guarantee a zero-advisory dependency tree.
Production-only dependencies are separately audited at moderate severity.

Install and rollback

Once the publishing workflow has completed, get LineLoom-1.4.2.vsix and its
LineLoom-1.4.2.sha256 from the
GitHub v1.4.2 release.
Verify the file digest and install it inside VS Code via
Extensions → … → Install from VSIX (not Visual Studio's VSIX Installer).
To return to the published stable, use the reviewed
GitHub v1.3.0 stable build.
The GitHub prerelease does not update either extension marketplace, the
independent CLI, or the publisher ID.