diff --git a/docs/plans/project-health-audit.rst b/docs/plans/project-health-audit.rst
new file mode 100644
index 0000000..a4f3bbc
--- /dev/null
+++ b/docs/plans/project-health-audit.rst
@@ -0,0 +1,164 @@
+Project Health Audit
+====================
+
+Feb 2026. Full codebase audit after combat-timing-fields landed.
+
+The architecture is solid. The problem is drift: things got built ahead of
+being cared about, so docs/tests/schemas are out of sync with actual code.
+Nothing is broken, but the project needs a cleanup pass before pushing
+forward.
+
+
+Combat: Doc/Schema Mismatch
+----------------------------
+
+The most concrete issue. Three things are wrong:
+
+1. **State machine docs are stale.** combat.rst describes
+   IDLE -> TELEGRAPH -> WINDOW -> RESOLVE (4 states). Code has
+   IDLE -> PENDING -> RESOLVE (3 states). TELEGRAPH and WINDOW got
+   collapsed into PENDING. The docs need to match reality.
+
+2. **Field names drifted.** Old schema used ``timing_window_ms`` for
+   everything. New schema split it: ``hit_time_ms`` for attacks (time to
+   impact), ``active_ms``/``recovery_ms`` for defenses (active window /
+   lockout). builder-manual.md and combat.rst still reference the old name.
+
+3. **No schema validation.** ``moves.py`` checks 3 required fields (name,
+   move_type, stamina_cost). You can define an attack with hit_time_ms=0
+   and it silently does nothing. Attacks should require hit_time_ms > 0,
+   defenses should require active_ms > 0.
+
+Other combat notes:
+
+- Stamina deduction is asymmetric: attacks deduct inside
+  ``encounter.attack()``, defenses deduct in the command handler before
+  calling ``encounter.defend()``. Works but confusing. Consider
+  standardizing.
+- Defense telegraphs are all empty strings. Intentional (reactive moves
+  don't broadcast intent) but undocumented.
+- Future features (stuns, combos, lethal, multi-combatant) are NOT
+  half-implemented. No stubs or dead code. Clean.
+
+Files to touch::
+
+    docs/how/combat.rst          - fix state machine, fix field names
+    docs/builder-manual.md       - fix TOML schema examples
+    src/mudlib/combat/moves.py   - add per-move-type validation
+
+
+Tests: Mixed Quality
+--------------------
+
+~1,448 test functions across 100+ files. The good tests are genuinely good
+(combat encounter, quetzal roundtrip, content loading). But there's bloat.
+
+**Trivial constructor tests (~50+).** Things like::
+
+    def test_thing_creation_minimal():
+        t = Thing(name="rock")
+        assert t.name == "rock"
+
+These verify Python dataclasses work. Delete them.
+
+**Over-mocking.** Many tests mock the writer then assert "was something
+written?" but not what. Color bugs, format bugs slip through. Loose
+assertions like ``assert "\033[" in result`` check some ANSI was emitted
+but not the right color.
+
+**Duplicated fixtures.** ``mock_writer`` is defined in 15+ test files
+instead of sharing from conftest. Same for ``mock_reader`` and zone helpers.
+
+**File sprawl.** 4 container test files, 3 portal files, 3 help files,
+4 zone files. These should consolidate into 1-2 per feature area.
+
+**Stub files (1 test each).** Seven files with a single test: test_corpse,
+test_npc_behavior, test_npc_integration, test_import, test_mobs,
+test_game_compatibility, test_two_way_portals. Either flesh out or delete.
+
+**Missing test categories:**
+
+- Error paths (bad input, wrong types, missing data)
+- Edge cases (boundary values, empty inputs, max values)
+- Concurrency (async code with no race condition tests)
+
+Plan::
+
+    1. Consolidate mock_writer/mock_reader into conftest
+    2. Delete trivial constructor/property tests
+    3. Merge fragmented test files (containers, portals, help, zones)
+    4. Decide on stub files: flesh out or remove
+    5. Strengthen loose assertions where practical
+
+
+Documentation: Strong Foundation, Gaps
+---------------------------------------
+
+**Current and good:** DREAMBOOK, architecture-plan, object-model,
+builder-manual, protocols, persistence, terrain-generation, IF docs,
+all lessons.
+
+**Stale:**
+
+- combat.rst (wrong state machine, wrong field names)
+- builder-manual.md (old timing_window_ms references)
+- roadmap.rst (phases 1-2 done but not marked)
+- IF research docs (viola/zvm/mojozork audits predate implementation)
+
+**Missing docs for existing systems:**
+
+- NPC/mob system (mobs.py, npc_behavior.py, npc_schedule.py, dialogue.py,
+  conversation.py, mob_ai.py - 6+ files, no implementation doc)
+- Thing/verb system (thing.py, things.py, verbs.py, verb_handlers.py)
+- Time/season/weather (gametime.py, seasons.py, weather.py, timeofday.py,
+  visibility.py)
+- Effects system (effects.py)
+- Targeting (targeting.py)
+- Loot/corpse (loot.py, corpse.py)
+- Crafting implementation (crafting.py - builder-manual has usage but no
+  internals doc)
+- Content loading pipeline (content.py)
+- Embedded z-machine (50+ files in zmachine/, no implementation guide)
+
+Plan: write docs as we touch each system, not all at once. Priority order
+matches what we're likely to work on next (combat first, then mobs/things).
+
+
+Architecture: Solid
+-------------------
+
+No action needed on architecture itself. Notes for awareness:
+
+- Module-level globals (players, mobs, active_encounters, command registry)
+  are appropriate for a game server. Tests clear them properly.
+- server.py is 676 lines and dense (login, shell loop, content loading) but
+  responsibilities are clear. Could split later if it grows.
+- Player dataclass has 29 fields. Organized but growing. Watch it.
+- Mode stack is string-based ("normal", "combat", "editor"). Works fine,
+  enum would be safer eventually.
+- No dead code in core engine. Z-machine has TODOs (expected).
+- No circular dependency issues. Lazy imports in commands/ handle cycles.
+
+
+Execution Order
+---------------
+
+Phase 1: Combat docs and schema (small, unblocks combat work)
+
+- Update combat.rst state machine and field names
+- Update builder-manual.md TOML examples
+- Add schema validation in moves.py
+- Tests for validation
+
+Phase 2: Test cleanup (can be gradual)
+
+- Consolidate fixtures into conftest
+- Delete trivial tests
+- Merge fragmented test files
+- Strengthen assertions
+
+Phase 3: Doc gaps (as-needed, per system)
+
+- Write how/ docs when touching each system
+- Mark completed roadmap phases
+- Consider archiving pre-implementation IF research docs