feat(qa): make GLM QA work end-to-end (scorer isolation + all-harness profile wiring)

qa/BEHAVIORAL_GATE_TAXONOMY.json

@@ -182,6 +182,12 @@
       "retest": "bash qa/run_duo.sh duo-retest",
       "hint": "A >=24-beat run (FELT_SHAPE_MIN_BEATS, strictly above the >=10 structural floor) covered 3+ acts but the arc never TURNED — no midpoint reversal AND/OR no late climax landed (a flat fetch-quest shape, not a felt setup->reversal->climax). DM adherence: drive a real midpoint reversal + a late climax (record_decision the turning beat, advance_act / mark_reversal / mark_climax, complete_quest the spine late). The felt_shape fields (reversal/climax/acts_engine_reached/acts_tag_reached) are produced by the acts-engine in server.py (NarrativeArc + advance_act/mark_reversal/mark_climax) and surfaced via qa/story_readout.py's felt_shape_from_state. WARN only (fatal=False); only armed at >=24 beats so shorter runs are unaffected."
     },
+    "dm_advanced_time": {
+      "category": "DM_ADHERENCE",
+      "likely_code_locations": ["qa/play_dm_duo.txt", "servers/engine/server.py"],
+      "retest": "bash qa/run_duo.sh duo-retest",
+      "hint": "The DM never issued a time-advance tool across the session — only the harness soft-tick moved the clock, so companion regard / camp / day-gated systems stay starved (#1024 unmasks the frozen-DM the soft-tick hid). DM adherence: call long_rest / advance_time / downtime when a beat should pass time. The time-advance tools live in server.py. WARN only (fatal=False)."
+    },
     "world_peopled": {
       "category": "DM_ADHERENCE",
       "likely_code_locations": ["qa/play_dm_duo.txt", "servers/engine/npc.py", "servers/engine/server.py"],

qa/run_combat_sprint.sh

@@ -17,6 +17,11 @@ cd "$ROOT" || exit 1
 
 RUN="${1:-cs-$(date +%H%M%S)}"
 WORLDOS_DM_MODEL="$(worldos_env DM_MODEL opus)"
+# GLM-only settings profile (no-op for Claude). Sourced after model vars resolve, before any
+# timeout/budget/retry knob is consumed. See qa/glm_profile.sh.
+# shellcheck source=glm_profile.sh
+. "$ROOT/qa/glm_profile.sh"
+worldos_apply_glm_profile
 SCORE_SCRIPT="$(worldos_env SCORE_SCRIPT qa/score.sh)"
 # Combat runs the whole multi-round fight on ONE budget (pre-seeded, no cold-open). An Opus combat
 # costs ~5x a Sonnet one, so the Sonnet-tuned $1.50 cap cut it off mid-fight (observed 2026-06-06:

qa/run_party.sh

@@ -61,6 +61,11 @@ SCORE_SCRIPT="$(worldos_env SCORE_SCRIPT qa/score.sh)"
 # Opus needs more than the Sonnet-tuned $0.80 per-call cap (the DM cold-open alone is ~$2.4); floor it
 # for an Opus DM so the cold-open lands. CAP, not spend; the Sonnet companion facade spends far less.
 case "$WORLDOS_DM_MODEL" in *opus*) if awk "BEGIN{exit !($BUDGET < 4.0)}"; then BUDGET=4.00; fi ;; esac
+# GLM-only settings profile (no-op for Claude). Sourced after model vars resolve, before any
+# timeout/budget/retry knob is consumed. See qa/glm_profile.sh.
+# shellcheck source=glm_profile.sh
+. "$ROOT/qa/glm_profile.sh"
+worldos_apply_glm_profile
 T="qa/transcripts"; STATE_DIR="$ROOT/qa/state/$RUN"
 mkdir -p "$T" "$STATE_DIR"; rm -rf "$STATE_DIR/campaigns" 2>/dev/null
 

qa/score.sh

@@ -84,7 +84,15 @@ while [ "$attempt" -lt 3 ]; do
   #
   # --json-schema was found to suppress the result text in this CLI; we rely on the
   # JSON-only instruction in the prompt and strip any stray code fences.
-  printf '%s' "$INPUT" | claude -p \
+  # GLM ISOLATION: the scorer is canonical Claude infrastructure (pinned sonnet, the gate
+  # baseline) and MUST run on clean Claude regardless of which model PLAYED the game. When the
+  # run is GLM, qa/glm_profile.sh exports ANTHROPIC_BASE_URL=z.ai + the GLM key + a fresh
+  # CLAUDE_CONFIG_DIR globally — which would route THIS scorer call to z.ai with a Claude model
+  # name (→ "Unknown Model") and skew/abort every score. Neutralize those vars for the scorer
+  # so it uses the default ~/.claude (Claude OAuth) + api.anthropic.com. On a normal Claude run
+  # these vars are unset, so `env -u …` is a NO-OP → byte-identical to today.
+  printf '%s' "$INPUT" | env -u ANTHROPIC_BASE_URL -u ANTHROPIC_API_KEY -u ANTHROPIC_AUTH_TOKEN \
+    -u API_TIMEOUT_MS -u CLAUDE_CONFIG_DIR claude -p \
     --model "$SCORER_MODEL" --permission-mode bypassPermissions \
     --max-budget-usd "$BUDGET" \
     --output-format json > "$RAW" 2> "$ERR"

qa/ui_playtest.sh

@@ -40,6 +40,17 @@ PLAYER_MODEL="$(worldos_env UIPT_PLAYER_MODEL sonnet)"
 # $1.50 cap trips error_max_budget_usd on the Opus cold-open (the PC never seats). Sonnet unchanged.
 case "$DM_MODEL" in *opus*) _uipt_dm_def=12.00 ;; *) _uipt_dm_def=1.50 ;; esac
 DM_BUDGET="$(worldos_env UIPT_DM_BUDGET "$_uipt_dm_def")"        # per DM turn (model-aware)
+# GLM-only settings profile (no-op for Claude). Sourced after model vars resolve, before any
+# timeout/budget/retry knob is consumed. This harness spawns the DM/player itself (it does NOT
+# delegate to run_duo.sh), so it must apply the profile directly. The profile keys off
+# WORLDOS_DM_MODEL/WORLDOS_ACTOR_MODEL, so export the role models under those names first (the
+# late WORLDOS_DM_MODEL="$DM_MODEL" near dm_turn re-asserts the same value, idempotently).
+# See qa/glm_profile.sh.
+WORLDOS_DM_MODEL="$DM_MODEL"
+WORLDOS_ACTOR_MODEL="$PLAYER_MODEL"
+# shellcheck source=glm_profile.sh
+. "$ROOT/qa/glm_profile.sh"
+worldos_apply_glm_profile
 PERSONA_FILE="$ROOT/qa/play_player_browser_${PERSONA}.txt"
 
 [ -f "$PERSONA_FILE" ] || { echo "[uipt] no persona brief at $PERSONA_FILE" >&2; exit 2; }