settled-reach/docs/workshops/test-architecture/round-2-notes.md

title, description, type, status, workshop, agent, round, created

title	description	type	status	workshop	agent	round	created
QA Strategy & Test Architecture Workshop — Round 2 Notes	Qatux's synthesis notes from round 2 cross-review and integration across all agents	workshop	archived	test-architecture		2	2026-02-17

QA Strategy & Test Architecture Workshop — Round 2 Notes

Workshop: QA Strategy & Test Architecture Round: 2 (Synthesis & Cross-Review) Date: 2026-02-17 Documenter: Qatux Participants: Tyre, Dudley, Stig, Hoshe, Justine, Gestalt, Ozzie

---

Executive Summary

Round 2 produced implementation-ready specifications across all 5 tracks. The lead's overrule on WS-D5 (separate test client binary instead of server-side text renderer) was accepted constructively by Tyre, who designed the binary architecture at server/src/bin/test_client.rs. Dudley delivered concrete code changes for all 4 determinism fixes (one was already done — a self-correction from Round 1). The cross-review process identified real coverage gaps: Hoshe found ZERO direct test coverage for the pause guard (all 6 of Dudley's proposed tests are genuine gaps), and Tyre added 6 tests to Stig's 32 for a total of 38 client assertions. The anti-tedium suite is fully specified across Ozzie, Stig, Gestalt, and Dudley. Content cross-reference validation is specified with 8 concrete checks.

Key convergence: Hoshe and Justine independently designed compatible make pre-pr targets. Ozzie and Stig independently designed compatible checklist formats (with Ozzie identifying 4 additions). Tyre validates Hoshe's CI tiers with refinements. All cross-reviews confirmed prior proposals as feasible.

One disagreement: fixture staleness should be BLOCKER (Tyre) or WARNING (Justine). Tyre's argument is stronger — stale fixtures mean client tests run against outdated protocol data, making all passing results false positives.

---

Lead Decisions — Agent Responses

The lead confirmed/overruled the 5 workshop decisions from Round 1. Here is how each was received and incorporated.

R1 Decision	Lead Ruling	Agent Response
WS-D1: Commit to determinism	ACCEPTED	Dudley produced exact code changes for 4 fixes (see Section 2). Self-corrected Fix C (monologue ordering was already done).
WS-D2: Hybrid YAML + Rust inject	ACCEPTED	Dudley designed GauntletRoom constants module + room reset mechanism. Gestalt mapped 14 rooms to the layout.
WS-D3: Rust test macros	ACCEPTED	No further discussion needed. Assertion helpers referenced throughout all proposals.
WS-D4: ObserverSnapshot golden files	ACCEPTED	Justine designed JSON golden file format with field-by-field diff. Dudley confirmed multi-position golden files (one per room).
WS-D5: Text renderer	OVERRULED — separate test client binary	Tyre accepted: "The architectural purity argument wins." Designed server/src/bin/test_client.rs with full CLI, text format, Layer 3 integration. See Section 1.

Additional lead decisions incorporated:

• No Gitea Actions for now — manual make ci stays. All agents adapted: Hoshe and Justine designed make pre-pr as the developer-facing tool. Tyre validated the CI tier design as "future-ready for when the lead greenlights CI."
• Anti-tedium full suite approved. Ozzie, Stig, Gestalt, and Dudley each contributed implementation specs.
• --test-mode + --port 0 confirmed for Sprint 8. Dudley designed the server flag changes (Section 3).
• Cross-reference validation added to make validate-content. Hoshe specified 8 checks; Justine specified 7 checks (compatible sets).

---

1. Test Client Binary Architecture (Tyre)

The lead overruled the Round 1 consensus (WS-D5: server-side text renderer) in favor of a separate Rust binary. Tyre designed the architecture.

Location

server/src/bin/test_client.rs — a second binary target in the server crate.

# server/Cargo.toml
[[bin]]
name = "settled-reach-test-client"
path = "src/bin/test_client.rs"

Rationale: Shares bridge types (ObserverSnapshot, PlayerInput, read_framed, write_framed) with zero duplication. Rust compiles each binary independently — no server bloat. A separate crate would need to depend on the server crate anyway (or require premature extraction of a protocol crate).

CLI Interface

settled-reach-test-client [OPTIONS]

Connection:
  --connect <host:port>     Connect to running server (default: 127.0.0.1:9876)

Input:
  --replay <file.jsonl>     Send inputs from file (one JSON PlayerInput per line)
  --interactive             Read inputs from stdin (future)

Output:
  --text                    Render each snapshot as structured text to stdout
  --json                    Dump each snapshot as JSON (for golden files)
  --quiet                   No output (assertions only, for CI)

Assertions:
  --golden <file.json>      Compare final snapshot against golden file, exit 1 on diff
  --ticks <N>               Disconnect after N ticks (default: unlimited)

Text Output Format

=== Tick 42 | Player (15,10) facing East | Stance: Walk | TickRate: Full ===
Game time: Day 0, 04:12 (Morning)
Entities (5):
  npc:100     (18,10) Forward  rel:Neutral     vis:Visible
  npc:101     (20,10) Forward  rel:Unknown     vis:Remembered
  obj:200     (16,9)  Forward  rel:n/a         vis:Visible
Tiles: 31 visible
Interactions (2):
  npc:100 [Talk(1), ExamineNpc(2)] distance=3
Inventory: 2/9 [item:300(slot-0), item:301(slot-3)]
Monologue: "Something about this manifest doesn't add up."
===

Entities labeled as kind:entity_id (e.g., npc:100). No display names on wire — test assertions use Gauntlet constants module: assert_entity_visible(&snapshot, GUARD_1.wire_id, GUARD_1.position).

Text Renderer Lives in Library

server/src/bridge/text_renderer.rs — library code callable by the test client binary AND integration tests. The server binary never references it.

Ozzie's Enhanced Terminal Layout

Ozzie proposed a richer live-updating terminal display for human testers (crossterm-based, ANSI escape codes), adding:

• Sound events section (D-018 three-range model)
• Cognition section (cognitive delay progress)
• Checklist progress bar
• Hotkey reminders (F12: WRONG, Home: Hub, R: Reset)
• Session timer
• Symbols per visibility state: ● VISIBLE, ◐ REMEMBERED, ◌ FOGGED, ✕ BLOCKED, ⚡ RECOGNIZING

Missing ObserverSnapshot Fields (Ozzie)

4 fields needed for the full test client experience, gated behind --debug/test-mode:

1. blocked_entities: Vec<BlockedEntity> — entities not visible with blocking wall position
2. Entity display names (or resolve via Gauntlet constants — Tyre recommends the latter)
3. Fog layer counts per type (5-layer breakdown)
4. Vec<SoundEvent> — source position, type, range classification

Effort Estimate

~2-3 days total: binary scaffolding (0.5d), text renderer (0.5-1d), replay loading (0.5d), golden file comparison (0.5d), Layer 3 wiring (0.5d).

---

2. Determinism Fixes — Concrete Code Changes (Dudley)

Dudley produced exact code changes for all 4 fixes identified in Round 1. Total: ~40 lines changed across 4 files (revised up from Round 1's estimate of ~15 lines due to full context being included).

Fix A: BTreeSet for visible_ids + sort visible_tiles

Files: server/src/perception/query.rs, server/src/perception/observer/mod.rs

Changes:

• visible_positions: HashSet<(i32, i32)> → BTreeSet<(i32, i32)> in VisibilityGeometry
• visible_ids: HashSet<u64> → BTreeSet<u64> in filter_visible_entities
• Add visible_tiles.sort_by_key(|t| (t.x, t.y)) after collection in NaturalVision::compute_geometry()
• Update collect_remembered_entities signature to accept &BTreeSet references

sector_lookup HashMap stays — it's point-lookup only via .get(), never iterated.

Fix B: Sort visible entities in snapshot

File: server/src/perception/observer/mod.rs, in compute_observer_snapshot()

Add after collect_remembered_entities:

entities.sort_by_key(|e| e.entity_id);

Fix C: Pin monologue system ordering — ALREADY DONE

Round 1 self-correction: Dudley stated monologue systems lacked explicit ordering. After re-reading bridge/mod.rs:168-175, confirmed the ordering is already explicit via .after() constraints. No change needed.

Fix D: Sort movers in validate_movement

File: server/src/simulation/movement.rs

Collect movers into Vec, sort by entity.to_bits(), then process:

let mut mover_list: Vec<_> = movers.iter_mut().collect();
mover_list.sort_by_key(|(entity, _, _, _)| entity.to_bits());

Note: entity.to_bits() provides stable ordering within a single run. For cross-session determinism (save/load), would need registry.to_stable(entity) — deferred to Sprint 10+.

Determinism Fix Coverage (Hoshe cross-review)

Hoshe assessed that none of the 4 fixes have direct test coverage today. Each fix should ship with its own regression test:

• Fix A: Test observer with 2 equidistant contradicted NPCs, assert same one selected
• Fix A (tiles): Generate snapshot with random-order tiles, assert output sorted
• Fix B: Not identified separately — covered by golden file tests
• Fix D: Two entities at same distance to same tile, assert deterministic winner

---

3. Server --test-mode and --port 0 Design (Dudley)

--test-mode Flag

Aspect	Behavior
Content	Loads Gauntlet content pack (content/gauntlet/), falls back to proof room
Seed	Fixed seed 42
Stdout	Prints LISTENING:{port} after TCP bind, before accept
Shutdown	Exits after first client disconnect
Logging	Defaults to warn (override with RUST_LOG)

--port 0 Support

Uses existing TcpBridge::accept_on(listener) (already in tcp.rs:67). The bind/accept split was added for test race condition prevention — exactly what --port 0 needs. No TcpBridge changes required.

Port discovery protocol: server prints LISTENING:{port} to stdout, test client parses it.

---

4. Cross-Review Results

4.1 Hoshe validates Dudley's server proposals

Pause guard: All 6 of Dudley's proposed tests are genuine gaps. The pause guard has ZERO direct test coverage today. The existing test (process_input_pause_sets_paused) tests the Pause action side effect, not the guard itself. If someone deleted lines 102-105 (the guard), all existing tests would still pass.

Hoshe found 3 additional gaps:

• set_tick_rate_while_paused (P2) — SetTickRate(Half) while paused unconditionally sets rate, making paused() return false
• perception_mode_while_paused (P2) — currently no-op but should pass through guard
• interact_take_while_paused (P2) — documents expected behavior for Take during pause

EntityRegistry: 3 of 5 Dudley tests are genuine gaps. 2 are not applicable (concurrent access impossible in bevy, bulk performance is fine at 2000 entries). Hoshe found 2 additional gaps:

• unregister_unknown_entity_is_noop (P1) — code handles it but no test
• register_with_pre_existing_stable_id_component (P2) — component/registry divergence

Bridge deserialization: Hoshe recommends keeping batch-failure (current behavior) with a test documenting it: malformed_input_in_batch_rejects_entire_batch.

4.2 Tyre validates Stig's client proposals

All 32 tests approved. No redundancies. Tyre adds 6 more (38 total):

#	Test	Category	Rationale
33	Pending recognition blob rendering	Entity rendering	D-060 cognitive delay visual
34	Recognition transition animation	Entity lerp	Blob → full entity over ~0.3s
35	Tick rate HUD indicator	UI elements	Full/Half/Paused display
36	Inventory full visual state	UI elements	9/9 feedback
37	Sprint interaction suppression	UI elements	D-055 interaction buffer empty during sprint
38	Entity modulate for Remembered	Entity rendering	Remembered vs Visible visual difference

Priority ranking: P0 = monologue not lost on overwrite (Bug #5), camera static during pause (Bug #2). P1 = fog shader state (3 tests), entity lifecycle, pending recognition blob.

Refinements:

• Camera rapid snapshots test should verify convergence, not just "doesn't crash"
• Entity lifecycle test should verify node freed (not just hidden) — memory leak prevention

4.3 Tyre validates Hoshe's CI tiers

Approved with refinements:

• PR tier budget revised: <15 min (not <10 min) — clean cache server builds take 5-7 min
• Add content cross-reference validation to PR tier (<2s, negligible budget impact)
• Add content scaling stress test to Nightly tier
• Fixture staleness should be BLOCKER, not WARNING — stale fixtures make all client tests false positives

4.4 Dudley validates Hoshe's Layer 3 test

Feasible with 2 minor adjustments:

1. Use rmp_serde::to_vec (not to_vec_named) for input serialization — matches GDScript's encoding
2. Tolerate initial tick=0 snapshot before input is processed — server may send snapshot before reading client input

Hoshe's assertions on tick=0 snapshot are fine — they verify snapshot delivery, not input processing.

---

5. Resolved Open Questions

ID	Question	Answer	Answered By
OQ-01	Does client depend on visible_tiles ordering?	No. All consumers are position-indexed (Dict keyed by Vector2i, set_cell() idempotent). Fix #2 safe to ship. Two fixture tests reference visible_tiles[0] by index — flag for fixture regeneration.	Stig
OQ-02	Does rmp_serde accept int_16 for u64?	Yes. Traced through rmp-serde 1.3.1: Marker::I16 → visit_i16 → visit_i64 → u64::try_from. GDScript encoding 256 as int_16 deserializes correctly. Negative values correctly rejected.	Dudley
OQ-04	WalkabilityMap HashMap → BTreeMap?	No. Leave as HashMap. Point-lookup only, never iterated. HashMap is faster in the hot path. Document as known-safe usage.	Tyre
OQ-05	Fixture staleness: git diff robust enough?	Yes. git diff --exit-code detects any byte-level change. Content-addressed hashing adds complexity for zero additional safety. Fixture generation is deterministic (verified — no timestamps/random values).	Justine
OQ-06	= OQ-02 (rmp_serde int_16→u64)	See OQ-02	Dudley
OQ-07	Server --test-mode and --port 0?	Designed. See Section 3. Uses existing accept_on(). LISTENING:{port} signal to stdout.	Dudley
OQ-08	Room name from coordinates?	Gauntlet coordinate bounds constants in server/src/test_world/constants.rs. room_at(player_pos) returns room name. No server API needed.	Dudley
OQ-10	Fog byte value constants?	Yes, promote to named constants. VIS_HIDDEN=0, VIS_PERIPHERAL=180, VIS_FORWARD=255, EXP_UNEXPLORED=0, EXP_EXPLORED=128, EXP_VISIBLE=255. Replace magic numbers in fog_state.gd.	Stig
OQ-12	= OQ-04 (WalkabilityMap HashMap)	See OQ-04	Tyre

Partially Resolved

ID	Question	Status
OQ-03	Gitea Actions available?	Moot — lead confirmed no CI for now. Tyre notes: ~1 day effort when greenlighted.
OQ-05	Entity display_name on wire?	Deferred. Tyre: use kind:entity_id labels now. Add display_name: Option<String> when client needs name labels (Sprint 9-10 gameplay feature).
OQ-09	Text renderer from ObserverSnapshot?	Yes, confirmed. The text renderer formats ObserverSnapshot. Lives in server library, called by test client binary.
OQ-11	Client tests headless?	Unresolved. No agent tested this in Round 2. Remains a concern for CI.

Resolved Unresolved Architectural Questions (from Round 1)

ID	Question	Resolution
UQ-01	Bridge deserialization: skip-and-log vs batch failure?	Keep batch failure. Both sides are co-versioned (D-020). Add test documenting behavior: malformed_input_in_batch_rejects_entire_batch. (Hoshe)
UQ-02	Gauntlet geometry: YAML vs Rust?	Not directly addressed in Round 2. Dudley's code uses Rust constants for room bounds. Gestalt references YAML room definitions for checklists. The hybrid approach (WS-D2) remains — specific geometry format deferred to implementation.
UQ-03	Wall-blocking-LOS info?	Optional blocked_entities field gated behind --debug/test-mode. Ozzie specifies as a requirement for the test client's ✕ BLOCKED display. Feasibility question to Dudley in Round 3.
UQ-04	Fixture path fragility?	Addressed by make pre-pr fixture staleness check. Regenerate + diff catches stale fixtures regardless of path.

---

6. Content Cross-Reference Validation

Both Hoshe and Justine independently specified content validation extensions. Their proposals are compatible and complementary.

Hoshe's 8 Checks

Check	Severity	What	Target
1	ERROR	canonical_id uniqueness across all NPCs	npcs/*.yaml
2	ERROR	Relationship target resolves to defined NPC	npcs/*.yaml → relationships[].target
3	ERROR	District locations[] slug matches location file	district.yaml → locations/
4	ERROR	Dialogue pool location matches district	dialogue/**/*.yaml → location
5	ERROR	knowledge_grant.fact_id validity	dialogue/**/*.yaml → lines[]
6	ERROR	NPC triangle_membership matches triangle file	npcs/*.yaml
7	WARNING	npc_count matches actual NPC file count	district.yaml
8	ERROR	Dialogue line ID uniqueness within pool	dialogue/**/*.yaml → lines[].id

Justine's 7 Checks

Check	Severity	What
1	ERROR	NPC relationship targets resolve
2	ERROR	Triangle members resolve
3	ERROR	NPC triangle_membership matches triangles
4	ERROR	Dialogue pool location resolves
5	ERROR	Fact IDs resolve (absorb check-fact-ids)
6	ERROR	District location list matches files
7	WARNING	Bidirectional relationship consistency

Merged View

Hoshe and Justine agree on checks 1-6. Justine adds bidirectional relationship warnings (Check 7). Hoshe adds dialogue line ID uniqueness (Check 8) and npc_count accuracy (Check 7). Combined: 9 unique checks (7 ERROR, 2 WARNING).

Architecture Agreement

Both recommend extending tooling/validate-content (Python) with a second pass after schema validation. No Rust dependency — content authors validate without compiling the server. Phased rollout: NPC/triangle/district first, dialogue/monologue second, warnings third.

---

7. make pre-pr Target

Both Hoshe and Justine independently designed this target. Their proposals are compatible.

Agreed Chain

pre-pr
  ├── 1. lint-server + lint-client          (~15s)
  ├── 2. build-server + build-client        (~30-90s)
  ├── 3. test-server + test-client          (~15-30s)
  ├── 4. validate-content                   (~2-5s)
  ├── 5. check-fact-ids                     (~2s)
  └── 6. fixture staleness check            (~10-15s)

Total: ~90-180s (under 3 minutes for clean incremental build). Fast enough for every PR.

Branch-Specific Variants (Hoshe)

• make pre-pr-server — lint-server, build-server, test-server, fixtures
• make pre-pr-client — lint-client, build-client, test-client
• make pre-pr-content — validate-content, check-fact-ids

Fixture Staleness Check

fixtures-check: fixtures
	@if git diff --quiet client/tests/fixtures/; then \
		echo "Fixtures: up to date"; \
	else \
		echo "FIXTURES STALE"; exit 1; \
	fi

---

8. Anti-Tedium Suite — Full Specifications

All features approved by lead. Four agents contributed implementation specs.

8.1 Room Reset Trigger

Aspect	Spec	Source
Trigger	Player steps on ResetPlate tile + presses Interact (not automatic)	Gestalt, Ozzie agree
Server mechanism	RoomResetTrigger component, RoomSnapshots resource (tick-0 state per room), execute_room_reset system	Dudley
What resets	Entity positions, entity KG, player KG (room refs only), fog (room tiles only), inventory items from room, dialogue state	Gestalt
What does NOT reset	Other rooms, player position (stays on plate), session timer, other-room checklist progress	Ozzie, Gestalt
Edge case	Items carried from room returned to tick-0 position, terminal shows "Items returned: keycard → crate_1"	Ozzie
Client visual	Distinct tile type (reset_plate), amber outline, interaction verb "Reset Room", 0.15s amber flash + monologue "Systems recalibrated."	Stig
Debounce	10-tick cooldown prevents re-trigger while walking across plate	Dudley
Test mode only	RoomResetTrigger entities only added with --test-mode	Dudley

8.2 Hub Teleport

Aspect	Spec	Source
Hotkey	Home key	Ozzie, Stig agree
Wire format	PlayerAction::TeleportToHub	Gestalt
Server behavior	Move player entity to GAUNTLET.hub_center, clear dialogue/monologue/interaction buffer	Dudley, Ozzie
Does NOT affect	Room state, inventory, game time, knowledge graph	Ozzie, Gestalt
Client visual	Instant camera snap, 0.3s fade-to-black-and-back, no monologue (meta action)	Stig
Gauntlet-only	Server rejects TeleportToHub in non-Gauntlet maps	Gestalt

8.3 WRONG Button (F12)

Aspect	Spec	Source
MVP captures (Sprint 8)	ObserverSnapshot, tick + position, text render output, human description	Gestalt
Full captures (Sprint 9+)	+ input history (60 ticks), snapshot history (60 ticks), world digest, replay seed	Ozzie
Output directory	tests/bug-reports/gauntlet-{tick}-{timestamp}/	Gestalt, Ozzie
Bug report format	report.md (human-readable), snapshot.json, text_output.txt, description.txt	Ozzie
Client implementation	BugReportCapture autoload, F12 hotkey in _unhandled_input(), modal prompt, 6 data captures	Stig
Ring buffer	60 ticks (configurable via --history-buffer)	Ozzie

8.4 Room Timer + Personal Bests

Aspect	Spec	Source
Display	TIMER: 00:47 (PB: 00:38) in status bar/overlay	Ozzie
Start	Player enters room (crosses bounding box)	Ozzie
Reset	Room reset trigger resets timer	Ozzie
Persistence	tests/gauntlet-stats.json — local, not committed	Ozzie
Session summary	Printed on disconnect: rooms visited, coverage %, times, PBs, bug reports filed	Ozzie

8.5 Auto-Checklist Progress (Ozzie + Stig merged)

Ozzie identified 4 gaps in Stig's per-room checklist proposal and proposed a merged format:

Gap	Stig's Proposal	Ozzie's Addition
No machine-readable conditions	Human prose only	Add condition: field (structured data) for auto-tracking
No auto vs manual distinction	All items equal	Add type: auto/manual — auto-confirms from snapshot, manual requires tester input
No cross-room items	Per-room only	Add cross_room_checks.yaml at Gauntlet root
No failure guidance	Just "check X"	Add if_wrong: field with likely causes + file references

Merged checklist YAML format (Ozzie):

checks:
  - id: occ_01_hidden_not_visible
    description: "NPC behind wall is NOT visible"
    type: auto
    condition:
      player_near: [15, 10]
      entity: hidden-1
      expected: blocked
    if_wrong: |
      LOS leaking through wall. Check shadowcast.rs.

make checklist generates markdown from YAML (Stig's tooling/gen_checklist.py). Test client loads structured conditions for auto-tracking.

8.6 Stig's Client-Side Anti-Tedium UI

• Room reset: Tile type reset_plate in TileRenderer, interaction verb "Reset Room", amber flash
• Hub teleport: TELEPORT_HUB in InputMapper → Home key, fade transition
• WRONG button: New bug_report.gd autoload, 60-entry input_history ring buffer
• Progress overlay: Top-right panel with room name, run counter, timer, checklist progress bar. Only visible when gauntlet_mode == true.

8.7 Deferred: F3 Debug Overlay

Stig recommends deferring the F3 State Inspector Overlay. The WRONG button captures the same data on demand. F3 as a real-time overlay requires per-frame string formatting of the entire ObserverSnapshot — measurable performance cost. Ship WRONG button first, F3 if testers ask for it.

---

9. Gauntlet Room Coverage (Gestalt)

Final Room List: 14 Rooms

#	Room	Primary Systems	Source
1	Inventory Warehouse	Pickup, CarriedBy, 9-slot limit	Brief
2	Occlusion Corridor	LOS, shadowcasting, perception modes	Brief
3	Interaction Gallery	ObjectType verbs, sprint suppression	Brief
4	Crowd Plaza	Entity density, relationship colors, cognitive delay	Brief
5	Fog Theater	Fog transitions, peripheral dimming, exploration persistence	Brief
6	Dialogue Room	Trust tiers, contradiction, walk-away, monologue during dialogue	Brief
7	Pause Chamber	TickRate toggle, state transitions	Brief
8	Zone Gate	Zone transition (reserved, future contract)	Brief
9	Eavesdrop Alcove	ListeningFocus, zone ambient, Careful stance	Gestalt R1
10	Confrontation Stage	Cognitive vulnerability, audio dip	Gestalt R1
11	Sprint Gauntlet	Sprint suppression, anomaly survival, stance transitions	Gestalt R1
12	Sound Lab	Three-range sound, sound pings, recognition chime	Gestalt R1
13	Decay Observatory	Knowledge decay, stale state, fog degradation	Gestalt R1
14	Shift Change	Stress test: all systems at density	Gestalt R1

Coverage Matrix

Gestalt mapped 50+ system-to-room relationships across 6 pillars: Characters & Information, Perception, Movement & Interaction, Audio, Simulation & Architecture, Content Systems. Every system from confirmed decisions has at least one room exercising it.

Coverage gaps (3 minor, all addressable without new rooms):

1. Object-layer favorite colors (D-052) — add to Inventory Warehouse in v0.1.2+
2. POI navigation (D-013) — add to Hub as POI markers. v0.1 stretch.
3. Environmental neutrality (D-045) — assertion on Dialogue Room (pre/post confrontation CanvasModulate identical)

8 Transition Scenarios

All scenarios reference physically connected rooms via hub paths or cross-cuts:

#	Scenario	Path	Key Test
T1	Sprint Exit	Plaza → Occlusion Corridor	Buffer cleared during sprint, LOS recalculated
T2	Fog into Dialogue	Fog Theater → Dialogue Room	Fog state preserved during dialogue, monologue above dialogue box
T3	Full Inventory Interact	Inventory → Interaction Gallery	9/9 inventory, Take still offered server-side, client greys out
T4	Sprint into Interaction	Sprint Gauntlet → Interaction Gallery	Verbs repopulate within 1 tick after stance change
T5	Confrontation to Eavesdrop	Confrontation → Eavesdrop	Audio dip release + ListeningFocus activation don't conflict
T6	Pause Anywhere	Pause Chamber → Hub → any room	Pause-during-transition state corruption
T7	Sound across Fog	Sound Lab → Fog Theater	Sound propagation through walls, cognitive delay from sound
T8	Walk-away Sprint	Dialogue → Hub → Sprint Gauntlet	KG incompleteness recorded, sprint suppresses post-dialogue monologue

Top 10 Invariants for Sprint 8 (Gestalt)

Rank	ID	Invariant	Bug Match
1	INV-T04	Pause coherence	Bug #3
2	INV-T01	Deterministic replay	Bug #1 class
3	INV-T03	Snapshot delivery	Bug #1, #5
4	INV-T05	Input ordering	Bug #1
5	INV-S01	Player spawn reachable	Softlock prevention
6	INV-C03	StableId uniqueness	Corruption prevention
7	INV-C07	Dialogue pool non-empty	Player-facing failure
8	INV-T02	Tick budget	Bug #6 class
9	INV-P02	LOS symmetry	D-035 mandate
10	INV-S05	No entity inside geometry	Content scaling safety

Bug catalogue mapping: every Sprint 6-7 bug is now covered by at least one invariant + room + test type.

---

10. Encoding Asymmetry Tests (Hoshe)

Hoshe specified 4-direction cross-language testing for the encoding asymmetry between GDScript (int_16 for 256-32767) and Rust (uint_16):

Direction	What	Test Location	When
Rust → GDScript (fixture)	Snapshots with overlap-zone ticks	test_msgpack_boundaries.gd	Every PR
GDScript → Rust (fixture)	Inputs with overlap-zone ticks	serialization.rs	Every PR
Rust → GDScript (raw bytes)	Hand-crafted uint_16/uint_32 bytes	test_msgpack_boundaries.gd	Every commit
GDScript → Rust (raw bytes)	Hand-crafted int_16/int_32 bytes	serialization.rs	Every commit

New make fixtures-client target generates GDScript-encoded fixtures for Rust to verify. Reverse direction of existing make fixtures.

---

11. Performance & Golden File Tooling (Justine)

Performance Baseline

• tests/perf/baseline.json — committed, records median/min/max from 5 runs
• tooling/perf-measure — builds release, runs benchmarks, compares against baseline
• Thresholds: <15% = PASS, 15-30% = WARNING, >30% = FAIL
• Machine tag prevents meaningless cross-machine comparisons
• make perf-baseline (compare) and make perf-baseline-update (update)

Golden File Diff

• Rust test, not separate tool — golden file is an ObserverSnapshot, Rust code knows the structure
• JSON format with sorted keys, pretty-printed (serde_json::to_string_pretty)
• Field-by-field diff output on failure: changed fields, POSITION markers, added/removed entities
• make golden-diff (view diff) and make golden-update (regenerate)
• Why JSON not MessagePack: human-readable in git diff, sorted keys = deterministic output

---

12. Client Test Suite — Final Count

Category	Round 1 (Stig)	Round 2 Additions	Total
Camera system	7	—	7
Entity rendering	7	+2 (Tyre: blob, remembered modulate)	9
Z-layer ordering	4	—	4
Fog shader state	3	+1 (Stig: hidden state)	4
UI elements	8	+3 (Tyre: tick rate, inventory full, sprint suppression)	11
Entity lerp	3	+1 (Tyre: recognition transition)	4
Anti-tedium	—	+2 (Stig: bug report capture, progress hidden)	2
Total	32	+9 (6 Tyre + 3 Stig)	~38-41

Note: Stig counts 35 (32+3), Tyre counts 38 (32+6). Combined unique total depends on overlap — upper bound is 41.

---

13. Open Questions for Round 3

New Questions (raised in Round 2)

ID	From	To	Question
R2-OQ-01	Hoshe	Dudley	SetTickRate(Half) while paused — should this unpause? Current code sets rate unconditionally (input.rs:165-168). Intentional or bug?
R2-OQ-02	Hoshe	Dudley	Entity respawn + registry — when bevy recycles Entity index, does registry handle old StableId not being unregistered?
R2-OQ-03	Hoshe	Tyre	make pre-pr — should it include make content-ron (YAML→RON conversion)?
R2-OQ-04	Hoshe	Justine	Fixture staleness in make pre-pr — separate make pre-pr-full to keep basic pre-PR fast?
R2-OQ-05	Ozzie	Dudley	blocked_entities debug field on ObserverSnapshot — feasible? Cost per tick?
R2-OQ-06	Ozzie	Stig	Can Godot client render checklist progress overlay, or test-client-only?
R2-OQ-07	Ozzie	Tyre	Test client binary location: server/src/bin/ or separate tools/ crate?
R2-OQ-08	Ozzie	Gestalt	Cross-room transition scenarios: where in YAML hierarchy? Own checklist section?
R2-OQ-09	Gestalt	All	Room ordering in Gauntlet YAML — canonical ordering affects entity StableId assignment.
R2-OQ-10	Gestalt	All	Per-room reset sufficient, or need "full server restart" command?
R2-OQ-11	Gestalt	All	4 new cross-cuts — too many? Consolidate Sound Lab into Occlusion Corridor sub-area?

Remaining Unresolved from Round 1

ID	Question	Status
OQ-11	Client tests headless stability	Unresolved — no agent tested in Round 2

---

14. Points of Agreement (consensus)

1. Test client binary is the right call. Tyre accepted the overrule and designed it properly. The binary shares types, exercises real TCP, and enables Layer 3 testing.
2. Determinism fixes are small and well-understood. Dudley's code changes are concrete. Hoshe's coverage assessment confirms no existing tests break.
3. Pause guard has zero coverage. Hoshe independently confirmed Dudley's gap analysis. All 6 tests are genuine needs.
4. make pre-pr replaces CI discipline. Both Hoshe and Justine converged on the same chain (lint → build → test → validate → fixtures).
5. Content cross-reference validation extends Python validator. Both Hoshe and Justine agree: no Rust dependency, two-pass architecture (schema then cross-refs).
6. Anti-tedium is fully specified. Room reset (Interact-triggered), hub teleport (Home key), WRONG button (F12 + MVP captures). No dissent on any feature.
7. 38 client tests are architecturally sound. Tyre's cross-review found no redundancies in Stig's 32 and added 6 meaningful tests.
8. Bug catalogue fully covered. Gestalt's invariant mapping confirms every Sprint 6-7 bug class has a test + room + invariant.
9. Checklist format merges Stig + Ozzie proposals. YAML with structured conditions for auto-tracking AND human prose for markdown generation.

15. Points of Tension

1. Fixture staleness: BLOCKER vs WARNING. Tyre argues BLOCKER (stale fixtures = false positive client tests). Justine tagged it WARNING. Recommend resolving in Round 3 — Tyre's argument appears stronger.
2. F3 debug overlay: defer or implement. Ozzie wants it for the testing experience. Stig says defer due to per-frame performance cost. WRONG button captures same data on demand. Lean toward deferral.
3. PR tier time budget: <10 min (Hoshe) vs <15 min (Tyre). Tyre accounts for clean-cache builds. Recommend <15 min as the budget (covers worst case).
4. Cross-room checklist location. Ozzie identifies a real gap: cross-room transition scenarios (Gestalt's T1-T8) don't fit in per-room YAML. Needs cross_room_checks.yaml or equivalent.
5. Number of Gauntlet cross-cuts. Gestalt proposes 4 new cross-cuts. Potential scope concern. Could consolidate Sound Lab into Occlusion Corridor sub-area.

---

16. Gaps Remaining

1. Test client binary doesn't exist yet. Tyre estimated 2-3 days. Depends on --test-mode + --port 0 (Dudley, Sprint 8).
2. Gauntlet content doesn't exist yet. Room YAML, entity placement, checklist definitions all need writing. No specific agent assigned.
3. Client test headless stability unverified. OQ-11 from Round 1 remains open. Must verify before making client tests a merge gate.
4. Cross-room checklist format unresolved. Ozzie identified the gap. Needs design in Round 3.
5. Gauntlet room ordering (entity spawn order → StableId assignment). Gestalt raised this. Affects determinism — canonical YAML ordering needed.
6. Sound events not in ObserverSnapshot. Ozzie requires Vec<SoundEvent> for full test client display. Not yet designed on server side.
7. Cognitive delay visual not yet implemented. Tyre's tests #33-34 (blob rendering, recognition transition) depend on D-060 implementation.