Function phase_has_human_only_checkpoint 

pub fn phase_has_human_only_checkpoint(
    project_root: &Path,
    phase: PhaseId,
) -> bool

Return true if any plan declared for phase DECLARES a checkpoint task that GSD will not auto-approve in any mode.

The match is anchored to a task element’s own opening tag, and that is the whole difference between this function and phase_has_blocking_human_checkpoint above: a marker qualifies only on a line that also opens a <task. A plan that merely discusses a marker — in a findings section, in a table, in a fenced example — has not declared one.

The concrete failures the anchoring prevents (F-14) were measured against this repository’s own plan files, not assumed: 34-04-PLAN.md:245 and 33-02-PLAN.md:109 each quote gate="blocking-human" inside a sentence while declaring no such task. Under an unanchored whole-file contains, preflight.rs’s unattended-launch check reads phases 33 and 34 as carrying a checkpoint no mode can approve and refuses an overnight run that was fine. A false refusal has no in-product recovery (D-09), so a false positive here is not a cosmetic defect. (F-14 itself names 35.1-03-PLAN.md as the instance; that file describes the markers only in English and never writes either literal, so it would not match an unanchored scan either. The finding stands; the example it cited does not.)

Known limit, pinned by human_only_checkpoint_still_matches_a_task_tag_ inside_a_fenced_example: the anchor is line-level and has no notion of markdown fences, so a plan documenting a COMPLETE example <task ...> tag inside a fenced block still matches. No plan file in this repository has that shape today.

phase_has_blocking_human_checkpoint is deliberately left alone, and the pair is not an accident. That function serves the plan-28-03 auto-decide route, where a looser, over-inclusive match fails SAFE — it routes more checkpoints to a human. Here an over-inclusive match fails toward refusing a launch. Different consequence, different predicate; widening the older one to serve both would silently change the behaviour its seven tests pin.

Returns false for a phase with no plan files at all. That is NOT the same fact as “plans exist and declare no such checkpoint”, and a caller that needs to tell them apart asks phase_plan_files separately rather than reading a third state out of one bit.

The CALLER owns root resolution, exactly as for phase_has_blocking_human_checkpoint: in worktree mode the phase’s plans live on the feature branch inside the worktree and are absent from the main checkout (999.76).