2.0.0a1 — the measured alpha: fourteen vetted mechanisms, graduated by evidence

PEP 440 pre-release: pip/uvx resolve this only with --pre (or an exact ==2.0.0a1 pin). Stable installs keep getting 1.14.5.

Fourteen externally-vetted mechanisms (ten Fractal-lineage, four jcode-lineage), built advisory-first under binding conditions, then graduated or deliberately kept dormant by a measured evidence matrix — never by taste. Per-mechanism verdicts with figures: D-077, D-078, D-079 in docs/DECISIONS.md; instruments committed beside the gameplan (matrix-p5-harness.py, matrix-p5-results.json).

Highlights (full detail in CHANGELOG):

• Honest endings: deferred-with-reason as a first-class close; laundering advisory; negative-space ("What I did not check") close-outs.
• Lifecycle detectors: stranded-state heal-on-proof (0 false positives across all matrix controls) and the interrupted-session backstop.
• Attention accounting: seen-vs-open receipts (drop-nothing, sidecar-gated), memory-gap detection at the moment of the gap, and the new cz_reinforce_lesson verb (strengthen instead of duplicate).
• Integrity: merge audit for canonical docs (3/3 seeded-fault detection, 0/4 false positives — advisory-silence lifted on evidence), enforcement ladder, transport-parity matrix.
• Fleet pattern productized: clauderizer-fleet ships as a wheel skill with measured guidance (D-079: quality tie solo-vs-fleet under independent adversarial verification; 1.57× wall-clock at ~1.7× compute; 0 collisions under the hub-and-spoke law).
• Dormant with figures: the per-call cz_state stamp (env-armed; DrvFs +86%/op priced default-on out) and budget wind-down (recording coverage measured first: 0.0 — dormant until post-publish telemetry).
• Hardening: mcp pinned >=1.2,<2 (H-31), stamps never ratchet backward (H-30), sanitizer removals echo to the writer (H-29), fresh repos scaffold GLOSSARY + ENFORCEMENT at every size.

Suite: 1330 → 1561 passing across the alpha; every new guard armed once (violation injected, red observed, reverted).

Known alpha caveat (by design, doctor tells you): clauderize init writes the portable MCP command (uvx --from clauderizer[mcp]), which resolves the latest stable — so a repo inited by the alpha will be served by 1.14.5 over MCP until you pin the server command to clauderizer[mcp]==2.0.0a1 (or use a local install). clauderize doctor detects and explains the skew explicitly. Alpha-only surfaces (e.g. cz_reinforce_lesson) need the pinned server; everything else works either way.