Name the scan-signal adapters of the Prompt Session

[derek-palmer]

From the 2026-06-09 Architecture Review. Recommendation strength: Worth exploring.

⚠️ Touches ADR-0001 territory but does not contradict it. ADR-0001 deliberately deferred unifying scan-state signals; this change keeps both signal behaviours exactly as they are and only gives each a name and a seam. The divergence stays intentional and becomes visible in the type system instead of living in two ad-hoc functions plus ADR prose.

Files

• src/codeforerunner/prompt_session.py
• src/codeforerunner/cli.py:17-23 (_scan_satisfied(): artifact ∨ FORERUNNER_SCAN_DONE env ∨ absent config)
• src/codeforerunner/mcp_server.py:88-124 (artifact ∨ in-process scan_called flag)

Problem

The Prompt Session owns the scan-first rule, but the contract a scan-signal source must satisfy is implicit: each adapter computes a bare scan_satisfied: bool its own way and injects it. A third adapter (a new delivery channel for Agent Onboarding, say) would have to rediscover both strategies from source and ADR prose, and could drift.

Solution

Introduce a small scan-signal protocol with two sealed implementations — CliScanSignal (artifact ∨ env ∨ absent-config) and McpScanSignal (artifact ∨ in-process flag). PromptSession accepts the adapter instead of a raw bool. Observable behaviour is identical for both adapters.

Benefits

• Seam: two adapters already exist, so the seam is real, not hypothetical.
• Tests: test_prompt_session.py parameterizes the gate rule over both signal adapters (plus a fake), instead of passing bare booleans.
• Locality: ADR-0001's rule/signal split becomes visible in code rather than folklore.