Dissent-span hiding channels the parser cannot close: comment shapes needing block context, plus non-comment raw HTML (`<script>`, `<template>`) — closure is an output-grammar rule

[Imbad0202]

Follow-up to #609 (PR #612), which closed the parser-side half of this and declared the rest as an explicit limit rather than leaving it undocumented.

The residual

A Phase 2 reviewer card that spells canonical dissent fields inside an HTML comment gets the fields struck from the canonical parse, so the seat is not credited with a dissent and keeps full Phase 1 trigger binding. #612 made that hold for every comment opener that sits at the start of a block, list and blockquote markers included (<!--, <!--, - <!--, 1. <!--, > <!--, > - <!--), and for openers that close and reopen on one line (<!-- an aside --> <!--).

Two shapes remain, and both do form a comment in CommonMark:

1. A marker following text on its own line.

   Reviewed the plan and stand by it. <!--
   dimension_id: D1
   rationale: plan was inadequate
   -->
   

2. A marker indented as a lazy paragraph continuation (four spaces starting a block is indented code, but after a paragraph line it is not).

   Reviewed the plan and stand by it.
       <!--
   dimension_id: D1
   rationale: plan was inadequate
   -->
   

Verified rather than argued: rendering either through MarkdownIt("commonmark"), parsing the HTML with html5lib and dropping comment nodes, the reader-visible text is Reviewed the plan and stand by it. — dimension_id: D1 is not in it. parse_dissent_dimensions nevertheless returns {'D1'}, so check_trigger_binding grants D1 its per-dimension exemption and a trigger that never appeared in the Phase 1 commitment passes.

The cost is stated at full price in #612 rather than discounted: this grants a trigger-binding exemption for a dissent the rendered page does not show, which is not what omitting the section would have done (omitting grants no exemption at all).

Why #612 did not close it in the parser

Closing it deterministically means reading a bare <!-- anywhere on a line as an opener. rationale: <nonempty explanation> permits any text, so a seat discussing an unclosed marker in its own explanation is a plausible card:

dimension_id: D1
rationale: the card left an unclosed <!-- marker in its own output
dimension_id: D2
rationale: the second plan was inadequate too

Under that rule the mid-line marker swallows D2's fields, they fail as unparsed occurrences, and the card aborts. Phase 2 permits no retry, so that abort takes the DA and the synthesizer with it — the exact failure #609 exists to remove, and the failure that fired on 1 of 2 panels on the 2026-07-27 fleet. That card is pinned as test_an_unbackticked_marker_in_a_rationale_hides_no_later_dissent.

Shape 2 additionally needs the walk to model the surrounding block context to tell lazy continuation from an indented example. Every mechanism added to that walk across the #612 review rounds produced a false abort of its own, which is why it stops where it does.

Proposed closure (output grammar, not parser)

Require the five Phase 2 seat prompts, the peer-review template and sprint_contract_protocol.md to state that comment syntax appearing in reviewer prose is written in inline code (`<!--`), the way the accepted card in test_a_canonical_rationale_may_mention_comment_syntax already writes it. With that rule delivered, the parser can treat any <!-- occurrence inside the dissent span as an opener, since a bare one is then out-of-grammar rather than plausible prose, and both residual shapes close.

Note this is a change to what reviewer cards may contain, so it needs the usual lockstep treatment: all five delivered Phase 2 subsections, the protocol mirror, the template, and role-scoped lint witnesses in check_role_scoped_contract.py.

An alternative worth weighing first, since it costs no grammar change: leave detection alone and stop treating a hidden claimed field as an abort at all, reading it as no dissent with the advisory [DISSENT-EMPTY-SECTION] line instead. That direction never grants an exemption either, and it removes an abort path rather than adding one; the tradeoff is that a malformed card becomes quieter. #612 kept the abort because "a claimed dissent aborts unless it is canonical" is the pre-existing invariant, not because the abort was measured to be better.

Where the current boundary is pinned

• test_a_comment_opened_after_prose_on_its_line_is_a_declared_limit (shape 1)
• test_an_indented_opener_continuing_a_paragraph_is_a_declared_limit (shape 2)
• test_an_unbackticked_marker_in_a_rationale_hides_no_later_dissent (the false abort that keeps the opener rule at block start)
• test_a_container_prefixed_opener_still_hides and test_comment_state_after_resolves_by_order (what fix(reviewer): #609 a dissent section that spells no field is no dissent #612 does close)

Each of the two limit tests says in its docstring that it documents an accepted miss and must be changed deliberately. Whoever picks this up should delete them in the same commit that lands the grammar rule, not before.

[Imbad0202]

Scope widened after measuring (2026-07-30)

The original write-up named only two comment shapes. Two later review passes on #612 showed the residual is larger and has a second, unrelated half, so the issue is restated here. Nothing below is a regression in #612 — every shape listed is credited on main too, and #612 strictly reduces the set.

Half 1: comment shapes that need block context

Measured against a real CommonMark render (markdown-it + html5lib, comment nodes dropped) over a 4032-shape grid — opener spelling (leading indentation x container marker x inner gap, spaces and tabs) x block context (block start / after a paragraph / inside a list item / inside a block quote) x closed-or-unclosed comment:

	hidden from the reader but credited	visible to the reader but refused
main	2572	0
#612 interim (guessed block state from non-blankness)	867	87
#612 as merged	1122	0

At a block start the checker and the renderer agree exactly in both directions. Every one of the 1122 surviving misses needs the opener to sit inside an already-open list item or block quote, or to be a lazy paragraph continuation. In those positions the container shifts the column at which a block starts, and changing marker type opens a fresh list, so a line that reads as indented text at an absolute column is a block opener relative to its container. Examples that reproduce: - an earlier note then 2. <!--, or <!--, or \t<!--; > an earlier note then > 2. <!-- or > <!--; and any paragraph line followed by <!-- with a later -->.

Resolving these means modeling container content columns, i.e. a container parser. #612 declined to grow one, on evidence: the interim version that approximated block state produced 87 false aborts, and a Phase 2 false abort kills the whole panel with no retry.

Half 2: hiding constructs that are not comments at all

The visibility model in #612 recognises exactly one construct, the HTML comment. Nothing in the scoring-seat path rejects other raw HTML, and markdown-it's commonmark preset passes it through verbatim. Canonical fields inside a wrapper whose contents no renderer displays are therefore credited as a genuine dissent and collect the trigger-binding exemption:

## Scoring Plan Dissent

<script>
dimension_id: D1
rationale: the phase 1 plan was inadequate here
</script>

No dissent: the Phase 1 plan holds.

The card renders normally, the dissent section positively asserts no dissent, and a trigger that never appeared in the Phase 1 commitment passes. <style> behaves the same; <template> and <div hidden> / style="display:none" are in the DOM but not displayed. <textarea> is not a vector, since its content is shown.

Extending the parser construct by construct is the wrong shape here, and #612's own history is the argument: a blacklist of wrappers was tried at the field-shape layer and was one wrapper behind at every round until it was inverted.

Proposed closure, one rule on the output side

Mirror what the DA path already does (check_panel_synthesis.py refuses HTML comments outright in DA reports) and state it in the reviewer output grammar: inside ## Scoring Plan Dissent, a reviewer card contains no raw HTML tag and no comment delimiter outside inline code. Then the parser can refuse any < … > tag or <!-- / --> token in that span as a [DISSENT-GRAMMAR: …] failure without guessing at block structure, which subsumes both halves and both declared comment limits at once.

The cost this pays, and the reason it belongs on the grammar side rather than in the parser: rationale: <nonempty explanation> permits any text today, so a seat discussing markup in its own explanation is a plausible card. Refusing bare markers without first telling seats to write them in inline code aborts that card, which is the failure #609 exists to remove. So the delivered prompts have to move first.

Needs the usual lockstep treatment: all five delivered Phase 2 subsections, the protocol mirror in sprint_contract_protocol.md, the peer-review template, and role-scoped lint witnesses in check_role_scoped_contract.py. The two declared-limit tests plus test_an_opener_inside_an_open_container_is_a_declared_limit should be deleted in the same commit that lands the grammar rule, not before.

Also worth deciding while here

#612 keeps the pre-existing invariant that a claimed dissent aborts unless canonical, which is why a field hidden in a fence or comment fails rather than being tolerated. Tolerating it instead (read as no dissent, advisory [DISSENT-EMPTY-SECTION] line) would never grant an exemption either, and would remove an abort path rather than add one. The tradeoff is that a malformed card becomes quieter. That choice was inherited, not measured, and this issue is the right place to settle it.