feat: bridge Codex v0.5–v0.6 runtime-parity-limits doc + changelog (4/4)

Phase 4 (final) of the Codex-Game-Studios v0.5–v0.6 bridge: document OpenCode runtime parity limits and record the full bridge in CHANGELOG/VERSION.

- README.md: new 'Runtime Parity Limits' section documenting OpenCode-specific enforcement limits honestly (advisory path rules, instruction/rule/hook-backed per-agent fences, model tiers as guidance, installer success != activation, workflow-guidance isolation). Also updates the deploy feature bullet (fail-closed preflight + --replace-modified + transactional rollback) and adds the .env* permission-layer note.
- CHANGELOG.md: v0.5.0 entry summarizing all four phases (secret protection, fail-closed installer, transactional rollback, advisory checks, parity-limits doc) and explicitly listing the Codex-only changes that were NOT ported.
- .opencode/VERSION: 0.4.2 -> 0.5.0; README badge + version refs updated.

Verified: audit.sh release passes (VERSION 0.5.0, CHANGELOG has v0.5.0 section); audit.sh all 0 errors; no stale 0.4.2 refs in README/VERSION. This completes the v0.5-v0.6 migration.

Author: Cipher-85

.opencode/VERSION

@@ -1 +1 @@
-0.4.2
+0.5.0

CHANGELOG.md

@@ -1,5 +1,63 @@
 # Changelog
 
+## v0.5.0 - 2026-07-12
+
+Bridged Codex Game Studios v0.5–v0.6 (upstream commit 259cff8) into the
+OpenCode-native port. Adopted the portable behavioral improvements only;
+Codex-platform internals (MultiAgent V2 role-activation proof, `.codex`
+installer/validators, CCGS framework frontmatter, skill delegation dialect)
+are intentionally not ported — OpenCode's Task/subagent model and `.opencode`
+structure make them inapplicable or already-equivalent.
+
+### Phase 1 — Secret protection + trust/activation messaging
+
+- Added `edit` and `glob` `*.env*` denies to `opencode.json` (the `edit` scope
+  covers write/patch), closing a gap where agents could create/overwrite
+  secrets via the write tool. Consolidated `read` to a single `*.env*`
+  catch-all so nested files and `.envrc` are covered; `*.env.example` stays
+  allowed for templates.
+- Install banner, README, and UPGRADING now state that installer success is
+  static-only and a new opencode session is required before hooks,
+  permissions, and agents are active.
+- `audit.sh config` now asserts an `edit` `*.env*` deny exists (regression guard).
+
+### Phase 2a — Installer/uninstaller fail-closed hardening
+
+- `coexistence.sh`: added `ccgs_state_validate` (exists/schema==2/valid
+  JSON/no path-traversal/no symlink) and `ccgs_state_owned_paths`.
+- `uninstall.sh`: fail-closed — removed the source-manifest fallback so
+  missing/invalid state aborts without removing files; dropped the redundant
+  emptiness-based AGENTS.md re-check; limited pruning to `.opencode/`.
+- `install.sh`: `--replace-modified` opt-in + Python preflight that aborts
+  before mutation on unowned collisions and locally-modified package files.
+- `audit.sh`: `run_install_safety` validator (5 static guards) wired into
+  `all` and exposed as `audit.sh install-safety`.
+
+### Phase 2b — Transactional deploy + rollback
+
+- `install.sh`: the cross-target deploy is now transactional — to-be-overwritten
+  files are snapshotted and created files recorded before mutation; a
+  mid-deploy failure restores modified files and removes created files, leaving
+  the target at its pre-deploy state. A `rollback-<timestamp>/` record is kept
+  for inspection.
+
+### Phase 3 — Advisory coexistence + smoke-headless checks
+
+- `audit.sh coexistence`: real install/uninstall matrix in a temp dir (fresh
+  install, collision abort, modified abort, `--replace-modified`, uninstall
+  missing-state, uninstall valid-state, transactional rollback).
+- `audit.sh smoke-headless`: command→skill graph integrity; model-driven smoke
+  deferred until a CI model runner exists.
+- `.github/workflows/release-check.yml`: `coexistence-advisory` and
+  `smoke-headless-advisory` jobs (`continue-on-error: true`); the existing
+  `validate` job remains the blocking gate.
+
+### Phase 4 — Runtime parity limits doc
+
+- README: new "Runtime Parity Limits" section documenting OpenCode-specific
+  enforcement limits honestly (advisory path rules, instruction-backed fences,
+  model tiers as guidance, installer success ≠ activation).
+
 ## v0.4.2 - 2026-07-10
 
 Bridged Codex Game Studios v0.4.5–v0.4.7 (bug lifecycle consolidation, handoff

README.md

@@ -13,7 +13,7 @@
   <a href=".opencode/skills"><img src="https://img.shields.io/badge/skills-77-green" alt="77 Skills"></a>
   <a href=".opencode/commands"><img src="https://img.shields.io/badge/commands-77-yellow" alt="77 Commands"></a>
   <a href=".opencode/hooks"><img src="https://img.shields.io/badge/hooks-12-orange" alt="12 Hooks"></a>
-  <a href=".opencode/VERSION"><img src="https://img.shields.io/badge/version-0.4.2-blue" alt="v0.4.2"></a>
+  <a href=".opencode/VERSION"><img src="https://img.shields.io/badge/version-0.5.0-blue" alt="v0.5.0"></a>
   <a href="https://opencode.ai"><img src="https://img.shields.io/badge/built%20for-OpenCode-f5f5f5" alt="Built for OpenCode"></a>
 </p>
 
@@ -69,7 +69,7 @@ first brainstorm to launch.
 
 ## Current Status
 
-Package version: `0.4.2` (see [`.opencode/VERSION`](.opencode/VERSION)).
+Package version: `0.5.0` (see [`.opencode/VERSION`](.opencode/VERSION)).
 
 This release includes:
 - Hard `/resume-from-handoff` lane-selection boundary: focus arguments only bias
@@ -107,9 +107,36 @@ This release includes:
   Code-Turn Discipline, Workflow Gates, File Lifecycle, Continuity Epilogue)
 - Model-agnostic installer with 3-tier model selection and hard validation
 - Coexistence detection for Claude Code, Codex, and mixed-runtime projects
-- Manifest-driven deploy with backup-before-overwrite and shared-path preservation
+- Manifest-driven deploy with fail-closed preflight, `--replace-modified`, and transactional rollback
 - Spawn-based plugin with payload capture for runtime verification
 - `'*': deny` permission model with `question`/`todowrite` allowed for all agents
+- `.env*` denied across `read`/`edit`/`glob`/`bash` so secrets are protected at the permission layer
+
+## Runtime Parity Limits
+
+This is a faithful semantic port of the upstream Claude Code Game Studios
+structure with documented runtime limits — not a claim that every upstream
+enforcement primitive is reproduced exactly in OpenCode:
+
+- **Path rules are advisory.** Root `AGENTS.md` routes edits to
+  `.opencode/rules/*.md`, but nested rule files are not auto-discovered by
+  OpenCode on every edit; agents are instructed to read them. Per-edit rule
+  triggering is instruction-backed, not a hard platform hook.
+- **Per-agent fences are instruction/rule/hook-backed.** Where upstream roles
+  deny specific tools, the OpenCode equivalent is enforced through agent
+  instructions, path rules, and the `ccgs-hooks.js` plugin rather than an exact
+  per-agent hard fence.
+- **Model tiers are guidance.** Per-agent `metadata.ccgs_tier` maps to a model
+  tier configured at install time; the OpenCode session model choice remains the
+  available runtime control.
+- **Installer success ≠ activation.** A successful install deploys files and
+  passes the static audit only. Trust the project and start a new `opencode`
+  session before hooks, rules, the permission profile, and agents are active.
+- **Prototype/worktree isolation** is explicit workflow guidance, not automatic
+  subagent isolation.
+
+The advisory `audit.sh coexistence` and `smoke-headless` commands surface
+regressions in these areas without blocking the build.
 
 ## Studio Hierarchy
 
@@ -415,7 +442,7 @@ opencode.json                        # Permissions, plugin ref, instruction file
   uninstall.sh                       # Coexistence-aware uninstaller
   audit.sh                           # Validation dispatcher
   release.sh                         # Version management + GitHub releases
-  VERSION                            # Package version (0.4.2)
+  VERSION                            # Package version (0.5.0)
 design/                              # GDDs, narrative docs (AGENTS.md + registry/)
 docs/                                # Technical docs, ADRs, engine reference
 production/                          # Sprint plans, milestones, session state