[johnnyhuy]

Unfinished tool calls now show Interrupted instead of Done in saved history, and active cards settle to Interrupted when the agent stops or its runtime exits. Partial output remains available, completed and failed calls retain their outcomes, and a late completion result clears the interrupted state.

Cards explain that a missing completion result does not establish whether changes were made. Markdown exports use the same status and count interrupted calls separately.

Part of #2.
Closes #47.

Verified locally with the pinned mise toolchain on macOS

• bun install --frozen-lockfile passed.
• bun run lint and bun run typecheck passed.
• bun run test:unit passed, 87 tests.
• bun run test:integration passed, 48 tests.
• mise run build passed.
• mise exec -- bun run test:e2e passed, 44 tests across five suites.
• The focused Electron stop scenario passed after the final spacing adjustment.
• GitHub CI, Electron E2E, CodeQL and Socket checks all passed on c2815bb before merge.

Proof

• .tmp/proof/e2e.log records the full Electron run.
• .tmp/proof/interrupted-e2e.log names the new stopped-tool scenario.
• .tmp/proof/interrupted-tool.png shows the expanded Interrupted card, warning and retained partial output. Visually inspected locally.
• Electron CI run also captures test-results/workbench/10-interrupted-tool.png in the existing pidex-e2e artefact.
• Named unit/component coverage is recorded in .tmp/proof/focused-unit.log and .tmp/proof/focused-integration.log for incomplete archives, empty completion results, late results, runtime exits, card wording and export counts.

The Electron suite uses deterministic Pi RPC fixtures. Live provider execution is not covered, and interruption does not imply that tool effects were rolled back. Scratch proof files are not committed.