Construct 3.0.0-alpha.28 — portable outcomes

Install the experimental alpha:

npm install -g @geraldmaron/construct@3.0.0-alpha.28

Construct now resolves portable research, decision, artifact, implementation and
review paths through the current host. The kernel retains context admission,
frozen execution contracts, evidence and the work ledger. No embedding model,
external vector service or inference runtime is installed or invoked by default.

Changes

• Portable intent captures outcome, audience, acceptance, constraints, assumptions
  and prior work. General chat delivery no longer requires file writing or code
  tests; verification follows the selected outcome.
• A rebuildable SQLite FTS5 index ranks admitted passages and project history.
  Bounded host query expansions and explicit relationships supplement exact lookup.
  Optional supplied vectors require configured model/dimensions and exact content
  digests. Missing vectors leave text retrieval operational.
• Source reads expose required schemas before reporting evidence. Directory sources
  retain bounded redacted supported text. Immutable snapshot citations preserve
  historical evidence; revoked access and changed current support remain enforced.
• Skill reports can bind material obligations to output sections and project
  evidence, with procedure references separate. Reports alone establish no benefit.
• Durable checkpoints preserve useful partial work. Delivery, deterministic checks,
  semantic review and person acceptance are reported separately. The shared reviewer
  contract is host-neutral; native adapter qualification remains explicit.

Upgrade

Run construct init --client=<host> and construct skill update, then restart
Construct sessions and inspect construct doctor. The operating skill is 3.1.0.
State remains format 5; the derived index requires no canonical-state migration.
Existing runs retain their frozen workflow and method bindings. Re-resolve an
outcome to use the new paths, preserving previous drafts and completed work.

Verification and limits

Lint, typecheck, the full suite (1,077 passed, one skipped), packaged isolated
smoke and six-host static conformance (97 passed, 19 untested) passed. A bounded
ordinary Codex decision trial selected the new path, retrieved prior context,
read local code, repaired a citation report and completed all three steps with a
grounded brief. It took 218 seconds and 13 Construct calls; efficiency needs work.

The canonical native isolation canary still fails. That single delivery diagnostic
does not qualify other outcomes or hosts, independent review, or causal skill
benefit. Earlier failed cohorts remain failures. Production source adapters,
executor qualification and the full live intake matrix remain open. Release uses
the existing finite experimental-alpha controls, not the qualified tier. Publish
only under alpha; latest remains on 2.1.1.

See the implementation record
and follow-up validation.

The release repeated the deterministic gate and three finite native held-text controls against this candidate. Correct arithmetic completed; incorrect arithmetic and the injected-source variant were rejected with held drafts preserved and no observed tools or fixture changes. Full qualification is not inferred.

Post-publication verification: official npm alpha is 3.0.0-alpha.28 and latest remains 2.1.1. The downloaded npm tarball has identical unpacked contents to the tested candidate, and the complete packaged consumer smoke passed against those published bytes. The verified tarball and SHA-256 checksum are attached.