feat(cli): stop accepting related_to in bulk (#508)

[jasonssdev]

First slice of #508 — the one that needs no prompt change. #508 stays open for the rest.

The hole this closes

PR #509 shipped --accept structure hours ago. It was applying, unreviewed, exactly the suggestions where the model had declined to claim anything.

related_to is not a wrong answer, and this is not a walk-back of the rubric. edge_typing.py is explicit that it "is defined as an ANSWER, not as a shrug", and the prompt says an honest related_to beats a guessed part_of, which "asserts something false about how the knowledge fits together, and anything reading this graph will believe it".

What sets it apart is narrower and uncontested: applying it adds no claim to the graph beyond the untyped link that was already there. That makes it the cheapest place to spend a human glance and — at 67% of accepted edges, measured on a real bundle (edge_typing.py:146) — also the largest. Note that corrects #385's "roughly 80%" from memory.

So --accept structure now applies every specific type without asking and routes this one to the operator. The exemption is scoped to the item, not the stage.

Non-TTY

On a pipe there is nothing to prompt on, so the exempted item is skipped, not prompted. Without that, the walk would reach typer.prompt with no terminal and die mid-run — the same failure the Identity non-TTY guard exists to prevent, and it exits 1 when mutated. An unattended run writes the confident suggestions and leaves the rest queued.

Why the numeric-confidence route did not ship

Investigating first changed the shape of the work. EdgeSuggestion and TierSuggestion carry no confidence, adding it is a prompt change — and no eval under evals/ scores either suggester. The cap harness measures extraction, a different pipeline. So there is nothing to A/B a prompt change against, and this project has already lost a prompt A/B once by adopting on intuition.

Building that harness is its own piece of work. #508 stays open for it; this slice delivers the behavior it was actually after, deterministically.

Metadata gets no equivalent: its three tiers are all substantive answers with no abstention token, and its volume is per concept type rather than per edge, so the prompt-count problem is Structure's almost entirely.

Verification

• Suite 4048 → 4050. ruff and mypy clean over the whole repo.
• Mutation-verified: re-accepting the abstention in bulk, and disabling the non-TTY branch, each fail their tests.

Disclosure: no review receipt

Same lens defect as PRs #503, #506, #507, #509 and #510.