[rabesss]

Fixes the P1 from the #90 council (round 5, Terra) that was downgraded to P2 and merged even though the command did not fail closed.

Problem

After the draft-feedback POST, _check_saved_comment checked the readback Html with regexes that accepted <p …> with any attributes, treated </p><p …> as a line break, and treated a raw \n as a line break. So Html: "<p hidden>hello</p>" (or style=/class=, a second paragraph, <p>x\ny</p>, a trailing <p>) with a matching Text reported saved-draft.

Change

• _html_text now runs a stdlib html.parser.HTMLParser (convert_charrefs=True) and accepts only the structure the command sends: whitespace, then one <p> with no attributes holding only text and attribute-free <br>, then </p>, then whitespace.
• Allowed reformatting: whitespace outside the paragraph, one line break directly next to <p>/<br>/</p>, entity spelling (&/&/&, " vs ", …), tag case, and <br>/<br/>/<br />. NBSP counts as a space and outer line breaks are ignored, as before.
• Anything else (any attribute, another element including <script>/<style>, comments, doctype, CDATA, PIs, a second or unclosed paragraph, </br>, an unfinished trailing tag, text after </p>, or any other raw line break) → mismatch → outcome: "unknown".
• I checked the matcher on CPython 3.11, 3.12, 3.13 and 3.14 against the edge cases (EOF handling of unfinished tags and comments, bogus comments). All four give the same results.
• Docs: updated the Readback 1 row and live-check step 5 in docs/assignment-feedback.md.

Tests

• Rejected even when Text matches: <p hidden>, <p style=…>, <p class=…>, <p id=…>, <br class=…>, span/div/img, <script> and <style> inside or after the paragraph, comments (also unfinished and bogus), doctype, CDATA, PI, a second paragraph (filled, empty or unclosed), <p/>, unclosed, no paragraph, a stray or double </p>, </br>, a trailing <b, text or an entity after </p>, a raw \n inside, two newlines after <br>.
• Accepted: <br> spellings, tag case, spaces inside tags, outer whitespace, line breaks next to tags (LF and CRLF), entity spellings, raw < text, NBSP, a dropped trailing break.
• _html_text(_html(t)) round-trips for a set of tricky texts.
• CLI: <p hidden> with matching Text → exit 1, outcome: unknown, never saved-draft.
• The old test that accepted paragraphs and attributes-nbsp is replaced; both shapes are now rejected.

Scope

Owner wrap-up: this PR only fixes the P1. The #91 items (cancel-document ids, Ctrl-C before the POST, published-warning ordering, "Nothing was sent." on pre-POST transport errors, broadening _send, the duplicate method check in write(), TOCTOU, live shape) and the round-5 P3s that this change doesn't already fix stay open on #91. The parser fixes two of those round-5 P3s: a raw \n counting as a line break, and an accepted trailing <p>.

Refs #91

Gate (head 68df29e)

• Round 1 (b1da724): Terra found three false saved-draft shapes: end-tag junk, dropped invalid charrefs, and ignored outer line breaks. The check now validates raw tokens: it checks each tag's spelling, slices the text between tags from the source, decodes it with a strict reference decoder, and compares line breaks exactly.
• Round 2 (68df29e): Opus high found no P0/P1, and so did Muse Spark xhigh. Terra xhigh raised one P1: zero-padded charrefs over 8 digits give unknown. That fails closed, so it was lowered to P2 and filed in assignment-feedback: follow-ups from #94 reviews (fail-closed edges and tests) #95 with the P3s.
• Matcher checked by hand on CPython 3.10.17, 3.11.12, 3.12.10, 3.13.3 and 3.14.8. Required CI checks pass. droid-review failed because Droid hit its usage limit; it isn't a required check.

[pullfrog]

This run was cancelled before it finished.

  ^{｜ View workflow run ｜ via Pullfrog}

[qodo-code-review]

ⓘ Qodo reviews are paused because your trial has ended. Ask your workspace admin to add credits to resume reviews. Manage billing

[factory-droid]

Droid encountered an error —— View job

Droid could not run because a usage limit was reached: You've reached your 5-hour Droid Core usage limit (resets in 3h 0min).
Reload Extra Usage credits or wait for your limits to reset.

---

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no actionable issue was found in the changed code.

Summary

This PR makes feedback readback stricter. _html_text now accepts one attribute-free <p> containing text and attribute-free <br> tags.

• Draft readback checks accept only the comment’s expected paragraph structure.

Scope acknowledgment: rabesss explicitly deferred the remaining #91 items: cancellation IDs, interrupts before the POST, published-warning ordering, pre-POST transport wording, broadening _send, the duplicate method check in write(), concurrent-write timing, and live response shapes. Other round-5 P3 items remain deferred unless this parser change already fixes them.

_{Reviews (1) · Last reviewed commit: "fix(instructor): fail closed on restruct..." · Reviewed by Greptile}

[factory-droid]

Droid encountered an error —— View job

Droid could not run because a usage limit was reached: You've reached your 5-hour Droid Core usage limit (resets in 2h 45min).
Reload Extra Usage credits or wait for your limits to reset.

---