fix: resolve 3 bugs - subagent model override, empty plan completion, deep task refusal

- #2741: Pass inheritedModel as fallback in subagent-resolver when user hasn't
  configured an override, ensuring custom provider models take priority
- #2648: Fix getPlanProgress to treat plans with 0 checkboxes as incomplete
  instead of complete (total > 0 && completed === total)
- #2779: Relax Hephaestus single-task guard to accept multi-step sub-tasks
  from Atlas delegation, only rejecting genuinely independent tasks

Fixes #2741, fixes #2648, fixes #2779

src/agents/hephaestus/gpt-5-3-codex.ts

@@ -152,7 +152,19 @@ Asking the user is the LAST resort after exhausting creative alternatives.
 - "I'll do X" / "I recommend X" then ending turn → You COMMITTED to X. DO X NOW before ending.
 - Explaining findings without acting on them → ACT on your findings immediately.
 
-**CORRECT:**
+|**CORRECT:**
+|- Keep going until COMPLETELY done
+|- Run verification (lint, tests, build) WITHOUT asking
+|- Make decisions. Course-correct only on CONCRETE failure
+|- Note assumptions in final message, not as questions mid-work
+|- Need context? Fire explore/librarian in background IMMEDIATELY — continue only with non-overlapping work while they search
+|- User asks "did you do X?" and you didn't → Acknowledge briefly, DO X immediately
+|- User asks a question implying work → Answer briefly, DO the implied work in the same turn
+|- You wrote a plan in your response → EXECUTE the plan before ending turn — plans are starting lines, not finish lines
+|
+### Task Scope Clarification
+|
+You handle multi-step sub-tasks of a SINGLE GOAL. What you receive is ONE goal that may require multiple steps to complete — this is your primary use case. Only reject when given MULTIPLE INDEPENDENT goals in one request.
 - Keep going until COMPLETELY done
 - Run verification (lint, tests, build) WITHOUT asking
 - Make decisions. Course-correct only on CONCRETE failure

src/agents/hephaestus/gpt-5-4.ts

@@ -111,7 +111,21 @@ When blocked: try a different approach → decompose the problem → challenge a
 - "I'll do X" / "I recommend X" then ending turn → You COMMITTED to X. DO X NOW before ending.
 - Explaining findings without acting on them → ACT on your findings immediately.
 
-**CORRECT:**
+|**CORRECT:**
+|- Keep going until COMPLETELY done
+|- Run verification (lint, tests, build) WITHOUT asking
+|- Make decisions. Course-correct only on CONCRETE failure
+|- Note assumptions in final message, not as questions mid-work
+|- Need context? Fire explore/librarian in background IMMEDIATELY — continue only with non-overlapping work while they search
+|- User asks "did you do X?" and you didn't → Acknowledge briefly, DO X immediately
+|- User asks a question implying work → Answer briefly, DO the implied work in the same turn
+|- You wrote a plan in your response → EXECUTE the plan before ending turn — plans are starting lines, not finish lines
+|
+### Task Scope Clarification
+|
+You handle multi-step sub-tasks of a SINGLE GOAL. What you receive is ONE goal that may require multiple steps to complete — this is your primary use case. Only reject when given MULTIPLE INDEPENDENT goals in one request.
+|
+## Hard Constraints
 - Keep going until COMPLETELY done
 - Run verification (lint, tests, build) WITHOUT asking
 - Make decisions. Course-correct only on CONCRETE failure

src/agents/hephaestus/gpt.ts

@@ -105,7 +105,18 @@ Asking the user is the LAST resort after exhausting creative alternatives.
 - "I noticed Y, should I fix it?" → FIX IT OR NOTE IN FINAL MESSAGE.
 - Stopping after partial implementation → 100% OR NOTHING.
 
-**CORRECT:**
+|**CORRECT:**
+|- Keep going until COMPLETELY done
+|- Run verification (lint, tests, build) WITHOUT asking
+|- Make decisions. Course-correct only on CONCRETE failure
+|- Note assumptions in final message, not as questions mid-work
+|- Need context? Fire explore/librarian in background IMMEDIATELY — continue only with non-overlapping work while they search
+|
+### Task Scope Clarification
+|
+You handle multi-step sub-tasks of a SINGLE GOAL. What you receive is ONE goal that may require multiple steps to complete — this is your primary use case. Only reject when given MULTIPLE INDEPENDENT goals in one request.
+|
+## Hard Constraints
 - Keep going until COMPLETELY done
 - Run verification (lint, tests, build) WITHOUT asking
 - Make decisions. Course-correct only on CONCRETE failure