Carry fail-open PreToolUse lapses to the gateway on the next PostToolUse (0.14.0)

[davidcrowe]

Why

An interactive-tier PreToolUse that cannot reach the gateway fails open (never-brick) and left its only record in ~/.acp/lapse.log. The gateway never learned the call ran ungoverned, so the console could not show it. Plugin half of davidcrowe/gatewaystack-connect#902; the gateway half is feat/pre-lapse-row there.

What

• failPostureOnOutage (interactive tier): in addition to the lapse.log line, append { at, tool, detail } to ~/.acp/lapse-pending/<session_id>.json (array, capped at the 20 most recent). Best-effort, never throws.
• handlePostToolUse: if a pending marker exists for this session, include pre_lapse: [...] in the /govern/tool-output body; delete the marker only after a 2xx. Existing 4s timeout untouched; a failed or unreachable report keeps the marker for the next call.
• UNGOVERNED systemMessage now says the lapse is queued for the session's next governed call (it previously said ACP had no record).
• Version bumped together in plugin.json, .claude-plugin/marketplace.json, and PLUGIN_VERSION → 0.14.0 (per gatewaystack-connect#888).

Tests

test/pre-lapse-report.test.mjs spawns the real hook against a closed port (outage) and a local stub gateway:

• fail-open still allows the call and writes the marker
• next PostToolUse carries pre_lapse, marker cleared on 2xx, following call is clean
• 503 / unreachable at PostToolUse keeps the marker
• repeated lapses append; cap at 20 holds
• one session's lapse never rides on another session's PostToolUse
• a corrupt marker is ignored and replaced

node --test: 105 passed, 0 failed · node scripts/validate-manifest.mjs: versions in sync (0.14.0).

Refs davidcrowe/gatewaystack-connect#902