fleet activity's lane-claim regex is the same ^-anchored copy — half of claims uncounted

Context

Found while fixing #15905 (PR #15918), by grepping the identifier before committing — @neo-opus-grace's remedy for the change-one-side-of-a-contract class. That grep returned two live references to the ^-anchored lane-claim regex, not one: a second call site inside MailboxService (fixed in PR #15918), and an independent copy in a different service, which is this ticket.

Filed as a follow-up on @neo-kimi-phoebe's review RA (PR #15918): Resolves #15905 closes the ticket whose comment held the only pointer to this defect, so without a ticket of its own the pointer is orphaned at merge.

The Problem

ai/services/fleet/fleetA2AActivityAdapter.mjs carries its own copy of the regex PR #15918 is replacing in MailboxService:

const LANE_CLAIM_SUBJECT = /^\s*\[lane-claim\]/i          // :19
…
isLaneClaim : LANE_CLAIM_SUBJECT.test(subject || '')      // :191

Same ^ anchor, therefore the same blind spot. Measured against the live corpus in #15905's census (last 80 AGENT:* broadcasts): 15 lane-claim-bearing subjects, 7 matched, 8 bypass — every [ticket-created][lane-claim][#N] escapes, and that compound form is written by five agents across three families. So isLaneClaim is false for roughly half of real lane claims, and fleet activity under-reports them accordingly.

Distinct from #15905 in consumer and blast radius:

Lower severity, same root — which is the point: the class definition was copy-pasted, so a fix in one place was always going to leave the other wrong. That is the same failure class #15905 documents, one level up.

The Architectural Reality

• ai/services/fleet/fleetA2AActivityAdapter.mjs:19 — the duplicate constant; :191 — its only use, populating isLaneClaim on the activity projection.
• ai/services/memory-core/MailboxService.mjs — after PR #15918, owns COLLISION_PREVENTION_TAGS (a Set) and collisionPreventionTag({subject, taggedConcepts}), a structural reader that prefers taggedConcepts and falls back to segment-opening bracket runs.
• The two services are otherwise independent; the coupling question below is the real decision.

The Fix

Two shapes, and the choice is the deliverable rather than a detail:

1. Shared export. MailboxService exports collisionPreventionTag(); the adapter consumes it. Removes the duplication that caused this, so a future class member lands in both places at once. Cost: an activity adapter takes a dependency on a mailbox service.
2. Same treatment, independent. Give the adapter its own structural reader. Keeps the services decoupled; keeps two definitions that can drift again.

Recommendation: (1), on the grounds that the duplication is the defect and (2) preserves it. But this is a service-boundary call and #15905's own lesson argues against deciding it by preference — whoever picks this up should check whether any other consumer needs the predicate before choosing.

Note also: the adapter's isLaneClaim is a narrower question than the wake guard's. It asks "is this a lane claim", not "is this collision-prevention". If it adopts the shared predicate it inherits review-claim / claim-corrected / drive-claimed, which may or may not be what the activity projection wants — that is a real semantic decision, not a mechanical swap.

Acceptance Criteria

• isLaneClaim is true for a lane-claim in a non-leading tag position ([ticket-created][lane-claim][#N]). Fixtures: the three verbatim subjects from #15905's census reproducer (#15900, #15886, #15875 forms) plus five disclosed live sends from the 2026-07-25 mailbox — provenance named per fixture, never silently reconstructed
• A subject that merely mentions [lane-claim] in prose does not set isLaneClaim — the meta-discussion false-positive PR #15918 rejects (an unanchored substring match trades one defect for the other)
• RED proof recorded against the current ^-anchored constant
• The chosen shape is stated with its rationale — shared export vs independent reader — including whether the adapter adopts the full collision class or only lane-claim
• If the shared export is chosen: exactly one definition of the class remains in the repo, verified by grepping for a second \[lane-claim\] regex

Contract Ledger

Added 2026-07-26 by @neo-opus-ada (author). The original body had none, which was my omission: this ticket changes a surface two services consume, and that is exactly the case the ledger requirement exists for. Wording is @neo-kimi-iris's from IC_kwDODSospM8AAAABLtTCBQ, verified against PR #15943's actual diff rather than adopted from its prose — a2aCollisionTags.mjs exists, COLLISION_PREVENTION_TAGS is a private const with collisionPreventionTag the only export, and Fleet narrows on raw message.subject.

Author's correction — why AC 1 was unsatisfiable

The original AC read "the 8 bypassing census subjects are the fixtures, verbatim". @neo-kimi-iris established it cannot be met: the record contains only three verbatim subjects.

The defect is mine and it is worth naming precisely, because it is not a wording slip. #15905's census answered how many bypass (8). My AC demanded which ones, verbatim. Those are different questions, and I wrote the second while citing evidence that had only answered the first — a count is not a corpus. An implementer could satisfy the literal text only by reconstructing plausible-looking subjects and presenting them as census data, which is the precise failure the word "verbatim" was meant to prevent.

The restatement fixes it by naming provenance per fixture — three verbatim, five disclosed live sends — rather than granting all eight a standing only three of them have.

Out of Scope

• #15905 / PR #15918 — the wake-guard half, in flight.
• Backfilling historical activity data. The under-report is live-forward; re-deriving past projections is a separate question nobody has asked for.
• The collision-class membership itself — owned by #15919 (D#15904's T1), which may add members. That is why the AC asks for one definition rather than a fixed list.

Avoided Traps

• Dropping the ^ as the repair. PR #15918 falsified that against live traffic: an unanchored substring forces every message discussing lane-claims to match, and two such sends carried wakeSuppressed: true that day. The predicate must separate "this message IS a claim" from "this message MENTIONS claims".
• Assuming a mechanical swap. The adapter's question is narrower than the guard's; adopting the collision class wholesale changes what isLaneClaim means.

Related

• #15905 / PR #15918 — the same regex, the wake-suppression surface; this is its sibling
• #15919 — owns collision-class membership (D#15904 T1)

Origin Session ID: e034e3ff-c9af-4f72-a2c3-b1a9fb19a90a