feat: per-workflow state isolation with project/ namespace (v10.2)

[volkz]

Problem

When two Claude Code windows work on the same repo simultaneously, they both write to the same flat activeContext.md — last writer wins, state gets clobbered. There is no isolation between parallel workflows on the same repo.

Solution

Adds a three-tier memory structure that eliminates cross-workflow state collision:

.cc10x/v10/
├── project/                    ← NEW: long-lived cross-workflow state
│   ├── activeContext.md        (architecture decisions, domain patterns, ongoing blockers)
│   ├── patterns.md
│   └── progress.md
├── workflows/
│   ├── <wf-id>.json            (existing hook artifact — unchanged)
│   ├── <wf-id>.events.jsonl    (existing — unchanged)
│   └── <wf-id>/                ← NEW: per-workflow isolated state
│       ├── activeContext.md    (current focus, active phase, in-flight tasks)
│       ├── patterns.md
│       └── progress.md
├── activeContext.md             (kept — backward-compat fallback)
├── patterns.md
└── progress.md

Two parallel Claude Code windows on the same repo each write exclusively to their own workflows/<wf-id>/ directory while sharing read-only access to project/ for durable project context.

Changes

scripts/cc10x_hooklib.py

• Adds project_state_dir() — returns .cc10x/v10/project/, creates it on demand
• Adds workflow_state_dir(workflow_id) — returns .cc10x/v10/workflows/<wf-id>/, creates it on demand

scripts/cc10x_pretooluse_guard.py

• Replaces the static protected-file set with _protected_memory_paths(), which dynamically covers all three tiers: root-flat, project/, and any existing workflows/<wf-id>/ directories
• Guard now audits/blocks writes to per-workflow memory files — parallel agents can no longer bypass it

scripts/cc10x_context_migration.py

• Adds _migrate_flat_to_project() — one-time intra-v10 migration that copies existing root-flat memory files into project/ on the first SessionStart after upgrading. Idempotent via a .project-namespace-migrated marker. Copies (not moves) to preserve the backward-compat fallback.
• Adds project and workflows to the source-discovery skip-list in _discover_sources() so the new namespace subdirectories are never treated as legacy migration sources

skills/cc10x-router/SKILL.md

• §2 Memory Load: updates the 4-step load to a 6-step two-tier load — always load project/ first, then load workflows/{wf}/ when a workflow_uuid is already known (resume path). Defines the merge precedence rule.
• §6 Workflow Task Graphs: adds step 4 — immediately after artifact creation, initialize the per-workflow state directory with mkdir -p .cc10x/v10/workflows/{workflow_uuid}
• §13 Memory Finalization: replaces the flat write list with an explicit write routing table. MEMORY_NOTES.patterns → project/patterns.md. MEMORY_NOTES.learnings + verification → workflows/{wf}/. Cross-workflow promotions clearly defined.

skills/session-memory/SKILL.md

• Updates Load-Bearing Boundaries to document the three-tier namespace
• Updates Memory Surfaces with per-tier descriptions and when to use each
• Updates Always Load with the ordered three-tier load sequence and merge rule
• Updates Red Flags to reference all three tier paths

Backward Compatibility

• Root-flat files are preserved (copy, not move) as a fallback layer
• Existing workflow artifact schema is unchanged — state_root still points at .cc10x/v10/
• The heading contract (required sections, stable anchors) is identical across all three tiers — no new templates needed
• Repos that have never generated a workflow_uuid continue working via the root-flat fallback until their first new workflow

Test Plan

• Fresh install: project/ created on first workflow; per-wf dir created after UUID generation in §6
• Existing install: cc10x_context_migration.py copies root-flat → project/ on SessionStart; .project-namespace-migrated marker prevents repeat runs
• Parallel windows: two Claude Code sessions on the same repo each write to distinct workflows/<wf>/ dirs; project/ is shared read context
• Pre-tool-use guard: write to project/activeContext.md from a write agent is audited/blocked
• Pre-tool-use guard: write to workflows/<wf>/activeContext.md from a write agent is audited/blocked
• Memory finalization: MEMORY_NOTES.patterns lands in project/patterns.md; workflow-local items stay in workflows/<wf>/
• _discover_sources() does not treat project/ or workflows/ as migration sources
• Repos with no workflow_uuid context continue to use root-flat fallback without errors