feat(phase5.3): guardrail — restrict resume tailoring to capable-tier models

User flagged: the tool runs Kimi/Qwen/Step/etc., not Claude Opus, so resume
quality varies by which model the round-robin lands on. Weaker models (Step,
Qwen-122b) drop roles and write sparse bullets.

Guardrail:
- config.py: added "tailor" flag. Capable tier (Kimi-K2.6, Qwen3.5-397b,
GPT-OSS-120b, DeepSeek-v4-Pro) = tailor:True. Weaker (Step-3.7-Flash,
Qwen3.5-122b ×2) = tailor:False (still used for fast bulk assessment).
- resume_customizer.customize_for_jobs: filters the tailoring cfg_pool to
tailor=True models only. Keeps parallelism across the strong models while
raising the quality floor. Weak models no longer generate resumes.

Note: the deterministic floor (weaving + backfill) already makes the SCORE
largely model-independent (~89% coverage regardless). This guardrail mainly
improves PROSE NATURALNESS by starting from a stronger first-pass.

No score regression: 8 JDs still 87-94 under calibrated scorer.

Action item recorded in memory: model-dependency-constraint.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

config.py

@@ -35,6 +35,7 @@ ASSESSMENT_MODELS = [
         "extra_body": {},
         "phase1":     True,
         "phase2":     True,   # ~5s per batch FAST
+        "tailor":     True,   # capable tier — strong at structured JSON rewrites
     },
     {
         "name":       "Step-3.7-Flash",
@@ -44,6 +45,7 @@ ASSESSMENT_MODELS = [
         "extra_body": {},
         "phase1":     True,
         "phase2":     True,   # ~8-35s FAST
+        "tailor":     False,  # weaker — drops roles / sparse bullets; not for tailoring
     },
     {
         "name":       "Qwen3.5-397b",
@@ -53,6 +55,7 @@ ASSESSMENT_MODELS = [
         "extra_body": {},
         "phase1":     True,
         "phase2":     True,   # ~9s FAST
+        "tailor":     True,   # capable tier — large model, good instruction following
     },
     {
         "name":       "Qwen3.5-122b-v2",
@@ -62,6 +65,7 @@ ASSESSMENT_MODELS = [
         "extra_body": {},
         "phase1":     True,
         "phase2":     True,   # ~12s FAST
+        "tailor":     False,  # smaller — keep for assessment, not tailoring
     },
     {
         "name":       "GPT-OSS-120b",
@@ -71,6 +75,7 @@ ASSESSMENT_MODELS = [
         "extra_body": {},
         "phase1":     True,
         "phase2":     True,   # ~11s FAST
+        "tailor":     True,   # capable tier
     },
     {
         "name":       "Qwen3.5-122b",
@@ -80,6 +85,7 @@ ASSESSMENT_MODELS = [
         "extra_body": {},
         "phase1":     True,
         "phase2":     True,   # ~40s OK
+        "tailor":     False,  # smaller
     },
     {
         "name":       "DeepSeek-v4-Pro",
@@ -89,6 +95,7 @@ ASSESSMENT_MODELS = [
         "extra_body": {"chat_template_kwargs": {"thinking": False}},
         "phase1":     True,
         "phase2":     True,   # ~42s OK
+        "tailor":     True,   # capable tier — strong reasoning
     },
     {
         "name":       "DeepSeek-v4-Flash",

src/resume_customizer.py

@@ -147,6 +147,17 @@ class ResumeCustomizer:
         print(f"  Template-only: {len(template_eligible)} jobs{Style.RESET_ALL}")
 
         cfg_pool = [c for c in (model_cfgs or []) if c and c.get("api_key")]
+        # GUARDRAIL (Phase 5): resume tailoring quality varies a lot by model.
+        # Restrict the tailoring pool to the "capable tier" (tailor=True) —
+        # Kimi/Qwen-397b/DeepSeek-Pro/GPT-OSS — and exclude weaker models
+        # (Step, Qwen-122b) that drop roles or write sparse bullets. The
+        # deterministic floor still backfills, but starting from a stronger
+        # model means more natural prose + higher first-pass coverage.
+        tailor_pool = [c for c in cfg_pool if c.get("tailor")]
+        if tailor_pool:
+            cfg_pool = tailor_pool
+            print(f"{Fore.CYAN}  Tailoring models (capable tier): "
+                  f"{', '.join(c.get('name','?') for c in cfg_pool)}{Style.RESET_ALL}")
         if not cfg_pool and self.fast_model_cfg:
             cfg_pool = [self.fast_model_cfg]
         n_workers = min(6, max(1, len(cfg_pool))) if cfg_pool else 1