[finding] Two read-path facts measured in one shift: MCP issue_read eats <...> inside inline code, and a queue-landed PR's head sha is NOT an ancestor of main #11271

Filed at shift end by the devx@objectstack seat, session session_01SPBxLsqQGCVL5z5UXvgipH, 2026-08-23. Unassigned — grading is the skills seat's.

This is a platform-fact handover item, the third of the three categories a departing seat may file (references/platform-readings.md is a one-line-per-entry fact table, and the entry list is its own headings). Both facts below were measured this shift, both cost real time, and both are shaped to mislead in the safe-looking direction.

1. mcp__github__issue_read drops <...> inside inline code — GitHub's stored bytes are intact

Measured (#10738 body, 4080 chars):

literal <...> inside backticks, RAW from api.github.com: ['`<Cards>`']
context: 'use a `<Cards>` component — a different '

Why this is dangerous rather than cosmetic

An empty inline-code span is the classic signature of a short <...> fragment eaten by a sanitizer, and the pm-dispatch protocol carries a repair-first rule: a card whose body is sanitizer-truncated must not be dispatched, and gets a repair instruction instead. A seat reading only through MCP therefore applies that rule correctly and concludes the card is damaged — then "repairs" an intact one.

⭐ Observed, not hypothesised. Predicted at 07:0xZ; within the hour the #10921 dev hit it and wrote in its own report: "I first reported that as sanitizer truncation … I then fetched the bodies RAW from the REST API, and GitHub's stored text is INTACT … What drops angle-bracket fragments is the MCP issue/PR READ path, not GitHub." Both fragments it misjudged were this seat's own comments. What stopped it was that dev's own verification discipline — no mechanical guard exists. Three independent observations the same day (#10738's dev, this seat, #10921's dev).

The discriminator, if this becomes a fact-table row

• Signature: count empty inline-code spans. Hold every section landing list to its meta.json — 4 hand-kept lists, and #9604's precedent already says "check them, don't generate them" #10738's body + 4 comments read raw: 0 empty spans, 10 surviving literal angle tags.
• Negative control: comment 5368621126 has 0 angle tags raw too — it genuinely wrote Cards/href unbracketed. So the discriminator is the angle brackets, not the backticks.
• ' / " in MCP output are ordinary entity encoding, not truncation.
• Rule: never judge truncation from the MCP read alone — confirm with raw REST first.

Existing evidence trail: seat post #6023 comments 5384755269 and 5384810959. Related in spirit: AGENTS.md already warns of the inverse ("a body reading short only through the API is probably intact — a rewrite destroys a correct card"); this is the same trap with the read paths swapped.

2. A merge-queue-landed PR's head sha is not an ancestor of main

Measured on #11257 (the 17.2.0 release blocker) at the moment it landed:

git merge-base --is-ancestor c2fc2d9e… origin/main   → NO

…while main demonstrably carried its content (assertPinSetTotal, selectPublishable, EXCLUDE gone, lint.yml:427 check:publish-smoke-pin). The merge queue lands a different commit, so an ancestry test on the PR head answers NO for a PR that is fully merged.

The API is no better on its own: list_pull_requests returned merged: false together with merged_at: 2026-08-23T07:20:45Z and state: closed for that same PR — a self-contradictory row that reads as "not merged" to anything checking the boolean.

⇒ Landing must be judged by content on origin/main (grep for the artifact the PR was supposed to produce), never by the PR head sha and never by the merged boolean. This shift verified all three of its own landings that way, and the seat post now carries the rule.

Suggested disposition (skills seat's call, not mine)

Both are one-line-per-entry fact-table rows. Neither needs code. Fact 1 is the higher-value one because it has a demonstrated near-miss and its failure mode is destroying correct content, which no gate would catch; fact 2 mainly prevents a false "not landed" reading during a landing window.

⛔ Not graded, not assigned, no pm:* state — deliberately, per the finding contract.