refactor(dream): remove learning pipeline + per-task skills with dynamic model loading

[dean0x]

Summary

This PR removes the Devflow self-learning pipeline end-to-end and restructures the Dream subsystem from a single sequential agent into N per-task agents with clean context and a hardcoded task→model map.

Delivered in two commits to allow atomic review of each concern:

• Phase A (cf853e5) — Learning pipeline removal end-to-end
• Phase B (98ffa97) — Per-task skills + dynamic model loading

---

Changes

Phase A: Learning Pipeline Removal

What was removed:

• scripts/hooks/eval-learning and eval-reinforce — SessionEnd learning batch accumulator and artifact reinforcement
• src/cli/commands/learn.ts — full devflow learn CLI command (5 subcommands)
• src/cli/hud/learning-counts.ts + components/learning-counts.ts — HUD learning counts display
• tests/learn.test.ts, tests/learning/hud-counts.test.ts, tests/learning/capacity-thresholds.test.ts, tests/learning/migration.test.ts — all learning-specific test files

What was updated:

• DreamConfig interface — now {memory, decisions, knowledge} (no learning field); coerceConfig silently drops legacy learning key
• scripts/hooks/dream-collect-tasks — learning) case is unconditional rm (delete orphaned markers on sight, not just when disabled)
• scripts/hooks/session-start-context — removed Section 1.75 (LEARNED BEHAVIORS) and all _SC2_LEARN_EN / _LEARNING_DREAM references
• src/cli/utils/migrations.ts — added purge-learning-pipeline-v1 (per-project) and purge-learning-global-v1 (global) migrations: clean up eval-learning, eval-reinforce installed hooks, orphaned learning log + config + state files, and the learning key from dream configs

Kept intentionally:

• src/cli/utils/learning-cleanup.ts — cleanSelfLearningArtifacts is called by the per-project migration
• src/cli/utils/observations.ts + observation-io.ts — decisions pipeline uses shared LearningObservation type and readObservations

Phase B: Per-Task Skills + Dynamic Model Loading

New skills (shared/skills/dream-{memory,decisions,knowledge,curation}/SKILL.md):

• Each carries the verbatim task procedure lifted from the former monolithic dream.md body — logic is byte-for-byte identical, only hosting changed (applies ADR-008: LLM judgment in skills, plumbing in scripts)
• allowed-tools: Read, Bash, Write, Edit, Glob, Grep — deliberate exception to the read-only skill default; these skills materialize artifacts (same posture as quality-gates and git)
• Lock hardening (reliability rule, AC-C6/AC-P4): replaced the give-up-fast mkdir || sleep 1 || exit 1 pattern in both the decisions skill (.reinforce.lock) and curation skill (.decisions.lock) with a bounded retry+backoff loop: 9 attempts, exponential backoff doubling from 1s capped at 8s, ~30s total. On exhaustion the task fails cleanly (leaves .processing for dream-recover) — never silently drops a write. No unbounded loops.
• Curation skill: acquires .decisions.lock EXACTLY ONCE across its read-modify-write; never calls decisions-append while holding it (avoids the deadlock documented in KNOWLEDGE.md)

Rewritten Dream agent (shared/agents/dream.md):

• Body is now plumbing-only: claim markers, heartbeat, multi-marker merge, then load the matching per-task skill via the Skill tool and follow it
• Supports the combined "decisions then curation" sequential spawn (Opus handles both, run decisions fully first, then curation)
• Adds four new dream skills to skills: frontmatter alongside existing devflow:apply-decisions and devflow:apply-feature-knowledge

Rewritten SessionStart spawn directive (scripts/hooks/session-start-context Section 2):

• Replaces single Agent(subagent_type="Dream") with per-task Agent() calls using a hardcoded task→model map: memory=haiku, knowledge=sonnet, decisions=opus, curation=opus
• decisions + curation co-pending → exactly ONE opus spawn whose prompt instructs running decisions skill then curation skill sequentially (prevents concurrent lock contention on .decisions.lock)
• Unknown task types are silently skipped (belt-and-suspenders — dream-collect-tasks should never emit them)
• bash -n clean; set -e intentionally absent (existing no-abort discipline, applies ADR-009/PF-008)

Registration:

• All four skills added to plugins/devflow-core-skills/.claude-plugin/plugin.json and src/cli/plugins.ts (core-skills is the correct home — skills install universally regardless of plugin selection)
• Bare names added to LEGACY_SKILL_NAMES for future cleanup migrations

---

Breaking Changes

None for end users. The devflow learn command is removed; users who call it will get a "command not found" error. The two new migrations clean up all installed artifacts automatically on devflow init.

---

Token Cost Characterization (AC-P1)

The new design spawns N per-task agents (up to 3: haiku for memory, sonnet for knowledge, opus for decisions+curation) instead of one sequential agent handling all tasks in a single context. Expected higher per-cycle token cost, justified by:

• Clean context per task: each agent starts fresh, reducing irrelevant context overhead and improving LLM focus
• Per-task model fit: memory (cheap haiku), knowledge refresh (capable sonnet), decisions/curation detection (high-quality opus) — the previous design ran everything at sonnet regardless
• Decisions quality: running decisions on Opus every deep session improves ADR/PF detection fidelity; user-accepted trade-off
• The concurrent execution model means wall-clock time is shorter even if raw token count increases

---

Reviewer Focus Areas

• shared/skills/dream-decisions/SKILL.md — bounded retry loop for .reinforce.lock; verify cap is 9 attempts, backoff doubles 1→2→4→8 (capped), no unbounded paths
• shared/skills/dream-curation/SKILL.md — verify lock acquired once, Edit calls happen between acquire/release, decisions-append not called while holding the lock
• scripts/hooks/session-start-context lines 183–245 — verify decisions+curation branch emits exactly ONE opus spawn; verify memory/knowledge branches emit their correct models; verify unknown types fall through to the "no known types" log path
• shared/agents/dream.md — confirm no learning remnants; confirm the "decisions then curation" combined prompt is unambiguous