23 tracked files cite three issue numbers that 404 (#611, #1854, #2295), so those obligations cannot be traced

Row: -

Needs an owner. Found while citing #611 as a blocker during A24 wave 5 (#2863);
#1854 and #2295 were found the same way during the adherence re-score (#2633).

FIVE cited issue numbers do not exist

gh api repos/mudler/vllm.cpp/issues/611   -> 404
gh api repos/mudler/vllm.cpp/issues/932   -> 404
gh api repos/mudler/vllm.cpp/issues/975   -> 404
gh api repos/mudler/vllm.cpp/issues/1854  -> 404
gh api repos/mudler/vllm.cpp/issues/2295  -> 404

#932 and #975 were added on 2026-09-06, found while looking for the next
dispatchable row. .agents/specs/ltx25-completion-scope.md:824 describes A15,
A16 and A17 as "one owed table with its issues already filed" and §4's rows
cite #975 for A15 and #932 for A16. Both are gone. The obligation itself is safe
— .agents/specs/ltx25-ic-lora.md lists them under ## Owed at :356 and
:357, which is what AGENTS.md requires — but a reader who chases the numbers
finds nothing, and the scope document asserts they are filed.

gh api user succeeds and every other issue queried in the same session resolves
normally, so this is not the API-lag or account-suspension signature. These
three are genuinely gone.

23 tracked files cite them, including .github/workflows/ci.yml,
scripts/ltx25-render-compare.py and several specs.

Why it matters

AGENTS.md makes an issue the unit of traceability: "An issue you do not fix in
the same flow has to say who owns it. It names an owning row in its Row: line,
or a spec lists it under ## Owed." A reference to a deleted issue satisfies
neither. A reader chasing one gets a 404 and cannot tell whether the obligation
was met, retired, or lost.

Two concrete instances found so far:

• #611 is cited as the blocker for the LTX-2.5 duration head's wiring. The
  obligation is in fact still owned — .agents/specs/ltx25-retire-dead-arms.md
## Owed says "A duration head is still not constructed" — so the substance is
  safe and only the pointer is dead.
• #1854 / #2295 are named as the hand-off in #2576's closing text, which was
  closed by #2633. That hand-off currently points at nothing; the live owner of
  the adherence gap is #2514, which is open.

So in both cases the work is owned by something real. The risk is that the next
reader cannot see that, and re-derives or re-files it.

What a fix needs to decide

• (a) Re-point each citation at the live owner (a spec ## Owed entry, or the
  open issue that actually carries it) and delete the dead numbers.
• (b) Re-create the three issues with their original content, if it can be
  recovered, so the existing citations resolve.
• (c) Add a checker that resolves every #NNNN a tracked file cites and fails
  on a 404 — note this needs a network call, which no checker in this repository
  currently makes, so it would have to be a CI-only gate.

(a) is smallest and (c) is the only one that stops it recurring. This needs a spec
if (c) is chosen, because it adds a checker with new semantics.

How they went missing is unverified

A mass gh issue edit run earlier in this repository's history caused an account
suspension, and the 404s that followed it were indistinguishable from deletions.
Whether these three were lost that way is not established here and should not
be assumed.

Start by reading the cited sections of .agents/specs/ltx25-completion-scope.md, .agents/specs/ltx25-ic-lora.md, and .agents/specs/ltx25-retire-dead-arms.md, then inspect the 23 tracked files and verify references with gh api. Done means each cited issue resolves or has a documented live owner; if a checker is chosen, its CI-only behavior and network semantics also need specification.