fix: recover from truncated resume session

[hyphen0009]

Description

Fixes an issue where ocr scan --resume fails with unexpected end of JSON input after a scan is interrupted with Ctrl+C.

The session uses buffered JSONL persistence, and an interrupted write can leave a truncated final record without a trailing newline. Previously, LoadResumeState treated this incomplete tail record as a fatal parsing error, preventing the scan from resuming.

This change allows resume loading to safely recover from an incomplete final record while keeping malformed records in the middle of the session file as errors.

The scan command also now handles the first Ctrl+C gracefully so checkpoints can be persisted and the session can be closed cleanly.

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change
• Refactoring
• Documentation update
• CI / Build / Tooling

How Has This Been Tested?

Added regression tests covering:

• Truncated final JSONL records
• Fragment-only session files
• Valid newline-less final records
• Mid-file corrupted records
• Review-session recovery
• Full abort → resume flow
• Concurrent checkpoint writes

Verification performed by the development agent:

• Session tests: PASS
• Full command test suite: PASS
• go build ./...: PASS
• go vet: PASS
• gofmt: clean

AI Assistance Disclosure

AI assistance was used during development to investigate and implement the fix and to help create the regression tests.

[CLAassistant]

All committers have signed the CLA.

[github-actions]

🔍 OpenCodeReview found 2 issue(s) in this PR.

• ✅ Successfully posted inline: 2 comment(s)

[github-actions]

The sessionID parameter is accepted but never used in this function. The function accesses the session ID through s.SessionID (via applyResumeLine) instead. Either remove the parameter or use it where needed (e.g., for error formatting). Keeping an unused parameter is confusing and will trigger linter warnings.

Suggestion:

Suggested change

	func (s *ResumeState) applyTrailingFragment(sessionID string, fragment []byte) error {
	func (s *ResumeState) applyTrailingFragment(fragment []byte) error {

[github-actions]

When the trailing fragment exceeds 256 bytes, it is silently truncated before being stored in RecoveredFragment. A user inspecting the fragment during diagnosis may not realize it was cut, leading to misleading forensic analysis. Consider appending a truncation indicator (e.g., "...") or storing the original length so the warning output makes the truncation obvious.

Suggestion:

Suggested change

	const maxFragment = 256
	if len(fragment) > maxFragment {
	fragment = fragment[:maxFragment]
	}
	const maxFragment = 256
	if len(fragment) > maxFragment {
	fragment = append(fragment[:maxFragment], []byte("...(truncated)")...)
	}

[lizhengfeng101]

Thanks for the fix! Heads-up: the core of this — tolerating a truncated final JSONL record in loadResumeState — was just landed by #1312 (merged), which fixes the same unexpected end of JSON input on --resume. Both branch from the same resume.go, so this will now conflict with main.

Could you rebase onto main and drop the parts already covered by #1312? A couple of things here still look like net-new value worth keeping:

• Graceful first Ctrl+C in scan_cmd.go / review_cmd.go / output.go so checkpoints persist and the session closes cleanly — out of scope for fix(session): tolerate truncated resume tail #1312.
• Recovering a valid-but-newline-less final record. fix(session): tolerate truncated resume tail #1312 unconditionally drops the EOF residue; your applyTrailingFragment parses it first and keeps it if valid. If that's intended, it'd be great as a small enhancement on top of fix(session): tolerate truncated resume tail #1312's logic (plus the Recovered / RecoveredFragment diagnostics), rather than a full rewrite of the loop.

So: rebase, drop the duplicated tail-truncation handling, and keep the Ctrl+C work (and the newline-less-record recovery if you want it) layered on top. Happy to re-review once it's rebased.