session: stop drawing the skills block as part of a message

[ZeroPoint95]

Head: 8501fad6e9663cdfefbf73e1a60bc24104d20ead — fix/skills-block-display → dev

What prints, and why

• Skills suited to this message: is per-turn skill selection — context for the model, spliced by attachTurnSkillsLocked onto the copy the provider reads.
• That copy lives in a.messages, which shapeEntries also walks to build every DisplayEntry — so Transcript, AttachReplay, and any reopen drew the block as though the person had typed it: top of new conversations, and after nearly every message whose words overlap the shelf.
• Skill-suitability block leaks into the visible transcript — keep it model-context only #1504 covered only the durable half: the journal and store always kept the typed words. The in-memory transcript is the half that was missed.

The fix: strip by injection provenance, never by text

• attachTurnSkillsLocked marks the one message it splices with the exact bytes it appended — a memory-only messagePresentation.SkillsBlock field riding the existing presentation index into shapeEntries.
• The mark is never journaled; a restored message carries no mark and needs none, because the journal keeps only the typed words.
• shapeEntries removes exactly those bytes from exactly that message. The model-bound copy is untouched.

What survives

• A person who pastes a whole skills block into a message of their own keeps every word — suffix matching (this PR's first cut, caught in review) could not tell the two apart; provenance can.
• Pinned cases: pasted complete block with no shelf; pasted block beside an injected one (only the injected copy comes off); marker/closing-line fragments; end-to-end Submit → Transcript shows the typed words while the provider copy keeps the block and the dim skills carried: notice is unchanged.
• Review follow-up: AttachReplay's own mid-turn slice is now exercised (it draws the running turn's opening message off a.messages[:floor]), and the provenance mark's pointer-identity assumption is pinned as a tripwire — recordUserLocked appends user.message without re-allocating Content (unlike recordPresentedAssistant, which deep-copies), so a deep copy on the user path would silently drop the mark and this test would go red.
• stripTurnSkillsBlock and plan's SkillsBlockConflictLine are deleted; plan/contract.go is back to dev's text.

Checks

• Build ✓, vet ✓, manual gate ✓; targeted skill/transcript/display/AttachReplay/provenance tests green. GitHub CI is the authoritative gate.
• Full internal/session suite on this branch fails 5 tests: TestHeldBeltRunStopsWithoutPreparingRepository, TestHeldRecoveryRetainsTheAcceptedCrew, TestHeldRunReopensOnTheSameAdmissionWithoutPreparingFiles, TestNewAdmissionReadsSettingsChangedBeforeItsCreation, TestStandingIsolationRecordsItsCopyBeforeWorkerInitialization. Failure classification: environmental, pre-existing — all five fail identically on a clean origin/dev checkout on this machine (macOS /var vs /private/var temp-path symlink in StandingIsolation, and held-run/admission timing timeouts under load in the rest). This PR touches none of them or their paths.

Known, out of scope (noted, not changed)

• questionconversation.go:117 builds the clarify child's model-facing context with displayEntries(a.snapshot()) — the index-less variant — so the injected block rides into that prompt. Pre-existing and model-facing (not a person-visible door), left as-is per review; noted here as an inconsistency with the display doors, which do strip.

—

Drafted with CodeAF · reviewed and owned by the author