# Phase 06.3 — Discussion Log **Session:** 2026-05-13 (cycle-4 closeout → cycle-5 scoping) **Mode:** /gsd-discuss-phase default **Phase status entry:** 06.2 CLOSED NOT APPROVED @ e4dcb3a, 06.3 inserted via /gsd-phase --insert (commit 3afeed9), then discuss. --- ## Gray Areas Presented (multiSelect) > Phase 06.3: cycle-4 gap-closure > Domain: Close cycle-4 UAT regressions on staging build 688d334. Re-run human UAT end-to-end. Capture CLI-08 milestone mp4 on PASS. **Options presented (6 candidates, trimmed to 4 by AskUserQuestion max):** - ☐ D-40/D-54 fix scope - ☐ D-45 attack pattern - ☐ D-51 evict crash diagnosis - ☐ D-55/D-56 scope decision (dropped to fit 4-max) - ☐ D-52/D-57 fold-or-split (dropped to fit 4-max — Claude defaulted) - ☐ GREEN gate uplift **Final presented (4):** D-40/D-54 fix scope, D-45 attack pattern, D-51 evict crash diagnosis, GREEN gate uplift. **Operator selected:** all 4. --- ## Area 1 — D-40/D-54 fix scope **Q1: How scope fix?** - Spike first — trace payload end-to-end ✓ (operator pick) - Direct fix — assume collision_polys missing - Split: collision-extent + outer-gray separately **Q2: Which rooms instrument?** - MVP-lobby only - All rooms in RoomRegistry - MVP-lobby + 1 sanity-check room ✓ (operator pick) **Q3: Spike telemetry surface?** - window.__rebno (same as 06.2-03) - Server log + client console - Both — window.__rebno + server log ✓ (operator pick) **Notes / refs surfaced:** - `packages/protocol/src/intents.ts:155-200` — newLayoutSchema does NOT define `collision_polys`; legacy shape (115-153) does. Schema gap is a real hypothesis. - `apps/client/src/render/RoomCollision.ts:23-60` — client derives walkable grid from `tiles[]`; sparse tiles[] coverage propagates to sparse collision. - 06.2-05 commit shipped tiles[] only (H1 fix path); broadcast did not extend to collision payload. --- ## Area 2 — D-45 attack pattern **Q1: Plan structure?** - Single combined investigation + fix plan - Spike + fix (flicker first, offset/depth second) - Three separate plans (offset / flicker / depth) ✓ (operator pick) **Q2: D-45 offset — derive correct formula from?** - Re-read legacy BNO Draw.dnd.json - Operator visual iteration - Both — legacy spec primary, visual confirm secondary - **Operator answer (freeform):** "we have historically referenced legacy. there must be a key difference between the font rendering tech here. here's what i can tell you for sure: the Y dist between the bottom of an 'a' and the top of the player's head is 10px in legacy, and 25px in the new client" (image attached confirming overshoot). **Q3: D-45 offset — which fix path? (added after operator measurement)** - Retune NAMETAG_OFFSET_Y constant (16 → 1) ✓ (operator pick) - Glyph-metric anchoring (canvas measureText) — deferred to Phase 7 - Switch to BitmapText (no padding) — deferred to Phase 7 **Q4: D-45 flicker root cause — likely path?** - Spike first — instrument flicker source ✓ (operator pick, with refinement) - Direct fix — floor to integer Y - Direct fix — pin to integer-floored sprite.y - **Operator refinement:** "it only happens on remote players, so the spike log would need to capture i.e. the first connected remote player, rather than the local player." **Q5: D-45 depth z-order — fix approach?** - setDepth(NAMETAG_DEPTH=10000) constant - Layer-based depth (Phaser Layer / Container) - Per-frame depth = sprite.y + 10000 - **Operator answer (freeform):** "there may be some things in game which overlap nameplates, but i'm not sure what they might be. please reference legacy and replicate how it's done there." **Notes / refs surfaced:** - `apps/client/src/render/Nameplate.ts:42` — NAMETAG_OFFSET_Y constant. - `apps/client/src/render/Nameplate.ts:98-100` — follow() formula. - `apps/client/public/atlas-mvp.json` — NaviStandD + NaviRunD both 36×48 (frame-bbox flicker hypothesis weak). - Phaser Text bbox includes line-height descender padding below visible glyph; legacy GM draw_text doesn't. This explains the 15-px overshoot operator measured. - Legacy depth_set refs: `extracted/client-5-8/objects/0042-player/events/Draw.dnd.json` + `extracted/client-5-8/objects/0000-server/events/Draw.dnd.json`. --- ## Area 3 — D-51 evict crash diagnosis **Q1: Which path?** - Spike-first — reproduce locally + trace broadcast ✓ (operator pick, with H1 disproof) - Direct fix on most-likely hypothesis (cookie/localStorage) - Stage spike+fix per hypothesis - **Operator refinement:** "hypothesis 1 disproven. they were different chrome profiles. -spike first" **Q2: Spike scope — what instrument?** - Full envelope — server outbound + client inbound + onJoin timing - Client-side only — onerror + handler trace - Playwright reproduction first; instrumentation if can't repro - **Operator delegated:** "agent decides using best judgment" → Agent picked **hybrid** (Playwright multi-tab fixture as primary + full envelope instrumentation in parallel). Fixture forces deterministic repro; envelope ensures root cause readable. **Q3: Fix contingency — if spike pinpoints H2 (seat race), is single-cycle fix OK?** - Single-cycle: spike + fix in 06.3 ✓ (operator pick) - Spike-and-defer — fix to 06.4 if non-trivial - Hard-close in 06.3 — disable dup-login on client too **Notes / refs surfaced:** - `apps/server/src/RebnoRoom.ts:409-495` — current eviction code looks correct; spike must reveal why BOTH tabs crash despite scoping. - H1 (shared cookie/localStorage) disproved by operator: different Chrome profiles. - Remaining hypotheses: H2 (seat-reservation race in Colyseus 0.17), H3 (Schema delete auto-broadcast side-effect on new-tab during onJoin), H4 (client-side force_reset handler error). - Colyseus 0.17 → 0.18 bump explicitly out-of-scope this cycle even if spike points there; would escalate to own phase. --- ## Area 4 — GREEN gate uplift **Q1: What MUST 06.3 add beyond unit tests? (multiSelect)** - Playwright e2e all GREEN on staging pre-merge - Operator visual confirmation pre-merge - E2e test reviews assertion-not-just-pass - Staging redeploy + smoke gate per fix - **Operator delegated:** "agent decides using best judgment" → Agent picked **all four**, weighted: (1)+(2)+(3) HARD gates, (4) SOFT (per-Wave-2-boundary, not per-fix). **Q2: Enforcement — how structurally bake in?** - PLAN.md per-plan gate checklist - Wave-3 hard verification with explicit checkboxes - Both — per-plan checklist + Wave-3 final gate - **Operator delegated:** "agent decides using best judgment" → Agent picked **option 3 (both)**. Defense-in-depth: per-plan catches mid-cycle drift; Wave-3 catches phase-level closure drift. **Notes / refs surfaced:** - Anti-recurrence pattern: cycle-3 `cli-08-nameplate-offset.e2e` PASS but caught nothing because expectation matched the (overshot) code. E2e assertion canonical-ref cite is the structural fix. - 06.1 ✗ → 06.2 ✗ — third NOT APPROVED would be a serious workflow signal. Phase 06.3 verification is treated as load-bearing. --- ## Discretion items (Claude defaults + operator overrides) **Defaults presented:** - D-55 (teleport in/out anim) — DEFER Phase 7 - D-56 (chat box readability) — INCLUDE 06.3 - D-52/D-57 — FOLD shared diagnostic spike - D-53 (spawn delay invisible) — small plan: publish + audit ensureLocal - CLI-08 mp4 capture — on PASS verdict **Operator overrides (freeform):** "Override D-55 — include in 06.3, roll in a plan to fix newly-identified bug: player sprite sometimes desyncs after player stops moving. so remote will look different from local --- and another one: in the screenshot, the only nameplate with the right font is 'uat_a' on uat_b's client. all of the 3 other nameplates are rendering with the old font (not fixedsys)" → Two NEW findings surfaced mid-discuss: - **D-58** — player sprite position desync after movement stops (remote view ≠ local view in steady state). Captured as reconciler-family. - **D-59** — nameplate font inconsistency. Only 1 of 4 nameplates renders Fixedsys. Hypothesis: WOFF2 document.fonts.ready race; Phaser Text snapshots font-availability at construction. **Q (follow-up): D-58 spike or fix?** - Fold into D-52/D-57 reconciler diagnostic spike ✓ (operator pick) - Separate spike — idle-state convergence pin-pointed - Direct fix — force re-sync on Stand state **Q (follow-up): D-59 spike or fix?** - Direct fix — gate Phaser boot on document.fonts.ready ✓ (operator pick) - Direct fix — re-render nameplates on font-load event - Spike first — confirm WOFF2 race --- ## Deferred Ideas - Colyseus 0.17 → 0.18 upgrade (only if D-51 spike requires it; escalates to own phase). - Glyph-metric-aware nameplate anchoring (alternative to D-45 Plan A constant retune; durable but heavier — Phase 7). - BitmapText / sprite-font glyphs (removes both font-metric padding + WOFF2 race — Phase 7). - Multi-room collision/payload generalization beyond MVP-lobby + sanity-check (Phase 7). - All Phase 06.1/06.2 cumulative defers (falls/abyss, ice tiles, Hexport, JokerShell, multi-floor, persistCharacter FK, prod provisioning). --- ## End-state CONTEXT.md committed; checkpoint cleaned. Next: `/gsd-plan-phase 06.3`.