PRISM v2.19.0 — Core-slimming refactor

MINOR, additive (no construct renumbering). Slice 5 — the terminal slice of
the dependency-ordered sequence: a projection-only, behavior-preserving
packaging refactor (Decision T). PRISM.md stays the canonical monolith; only the
projection and the load contract change.

Surface

• Core-slimming decouple (§A2 / Decision T). The always-loaded core sheds its
  phase-, seam-, and condition-scoped mechanics into four new on-demand
  bundles keyed to the role that needs them, leaving the execution-hot spine
  resident. Core ~24% leaner (5092 → 3895 lines).
  • reference/setup.md — Setup mechanics: the seven probes (P1–P7),
    three-layer readiness, the Setup-artifact procedures, strategy stability,
    onboarding (§6.1–6.6); Library-reference-at-Setup (§7.1); the Scope-Integrity
    Test + specialist-pass promotion (§7.3–7.4); the Claude-Project-as-Setup
    recommendation (§8.1). Trigger: no Master yet / operator initiates an
    engagement (P0).
  • reference/currency.md — point-refresh (§7.5), the Update session (§7.6),
    the session-open currency check (§18.3). Trigger: an Update session / the
    currency-check-at-open.
  • reference/continuity.md — migration handoff, failsafe recovery,
    defensive migration (§5.4–5.6), missing-handoff recovery (§14). Keeps a
    minimal in-core recovery floor resident (write-at-every-turn-close +
    recover-from-the-repo) so a session in trouble is never stranded. Trigger: a
    resume with a missing handoff, or M5 band 🟠/🔴.
  • reference/corpus-access.md — the corpus-access dispatch (§4.13).
    Trigger: the dispatch-builder classifies a pass as corpus-access.
• Fail-loud load triggers. Each gutted area leaves a small in-core trigger
  stub — a detector + pointer + a visible SP-4 self-check load line ("phase =
  Setup → setup.md ✓") — so operating without a required bundle is detectable,
  not a silent gap. The §3.7.6 phase→bundle manifest and the SKILL.md
  core-load gate route the loads; the v2.16 bundle-load integrity layer is the
  safety net.
• What stays in core: the execution-hot spine — Standing Principles,
  Monitors, the triple contract, the prompt-package engine, the dispatch
  lifecycle, the §6.7 closure gate, and the §7.2 lens schema (the dispatch
  interface).
• One doc-hygiene content correction folded in. The Appendix C tag-count
  summary is reconciled to its list-entry counts (structural 44→50, methodological
  12→18; totals 74 / 81), with a lockstep-invariant note so it can't silently
  re-drift; the shipped reference/provenance.md table now matches.

Behavior-preservation (the bar for a projection-only slice)

• Union of core + bundles is content-equivalent to the prior projection
  (line-accounting verified: core-uncovered + bundles + discarded == body; 0
  source lines lost, +75 generated stub lines). Every new bundle is reachable
  via its in-core trigger + the manifest + the SKILL.md load step; a miss is
  caught by the integrity layer. Projection idempotent. Lint 0 errors (named-refs
  on PRISM.md; cross-file on the projected tree).

Calibration / report back

• Do the in-core trigger stubs + SP-4 self-check reliably load the right bundle at
  the right moment across surfaces (deterministic skill-loader vs model-driven)?
• Does the leaner resident core measurably lower M5 context pressure on long
  Execution sessions?
• Is the in-core recovery floor enough to never strand a resuming session before
  it fetches continuity.md?

Provenance. Lens Library unchanged (v0.15). Snapshot PRISM_v2_19_0.md
(byte-identical at the v2.19.0 tag). The projection script (workshop
scripts/decompose/project_skill_archive.py) was reconciled onto origin's
v2.14.0 base + the 2.15→2.18 deltas (verified byte-for-byte against the committed
v2.18.0 tree) before adding the four phase-bundle regions. Decision T (full
conservative set, Setup-first, recovery-last with an in-core stub) and Decision S
(core-light routing) resolved.

---

Part of the PRISM v2.15.0 → v2.19.0 promotion arc. See RELEASING.md.