[HIGH][SECURITY] Corrupt or unreadable config silently disables enforcement while the banner still prints ENFORCING

[Aniketvish0]

Migrated from armoriq/conmap-auto#710.

Fix lands in: armoriq/armorclaude and armoriq/armorcodex. Filed here because conmap-auto is this session's tracker for CLI/SDK work.

Line refs are from origin/main: armorclaude @ 688bbc3 (AC), armorcodex @ 72d9e13 (AX).

Three variants.

(a) AC scripts/lib/policy.mjs:52-77 + scripts/lib/fs-store.mjs:12-16 — a SyntaxError parsing policy.json returns an all-allow default. It also disables the intent gate, because the resulting permissive policy is detected as "frictionless" (engine.mjs:84-91, 875-876). One corrupt byte removes both policy denial and the plan gate, with no signal.

(b) AX plugins/armorcodex/scripts/lib/policy.mjs:237 + fs-store.mjs:9-19 — same chain, terminating in return {allowed:true}. handleSessionStart (AX engine.mjs:162-167) still prints ArmorCodex active (ENFORCING…) because it reads only config.mode.

(c) AC config.mjs:90-98 / AX config.mjs:75 — an unreadable credentials file is swallowed, leaving apiKey:"", which silently disables audit, crypto verification, observability and intent-minting (AX engine.mjs:611-614, config.mjs:106,165-168) while the same banner prints.

Ask

• Fail closed on unparseable policy.
• Make the banner reflect the enforcement state actually achieved, not the configured mode.
• Log all three at error level.

[Aniketvish0]

Filed in both CLI plugin repos — the defect has a distinct instance in each.
Companion: armoriq/armorClaude#111