backlog-digest links PRs by closing keyword only, so a 'Part of #N' PR is invisible — #409 reported In Progress with an APPROVED PR open

[bess-product-owner]

Problem

backlog-digest.sh links a PR to its issue only through a closing keyword:

backlog-digest.sh:80   scan("(?i)(?:fixes|closes|resolves) #([0-9]+)")
backlog-digest.sh:168  test("(?i)(fixes|closes|resolves) #\($n)\\b")

But this project's convention forbids that keyword on any PR that is not the
final one for an issue — an intermediate or beta PR must not auto-close the
reporter's issue. The two rules are in direct conflict: following the
convention makes the work invisible to the board.

Measured — #409 / #490

PR #490 says, in its own body:

First of two PRs tracking #409 … Issue #409 stays open until that second PR
lands.
Part of #409 — this PR covers PredictionSnapshotStore only.

Three references to #409, no closing keyword, exactly as the convention
requires. The digest therefore resolves $pr == null for #409 and falls
through to:

backlog-digest.sh:339   if $pr != null then "In Review"
backlog-digest.sh:342   elif $wt_live  then "In Progress"

So issue #409 derives In Progress, and this pass emitted both:

##409 move_card           card is In Review but the evidence says In Progress
##409 resume_implementation   worktree on disk, unlocked, no live session

Both are wrong. PR #490 is out of draft, carries two APPROVED reviews, and
merges cleanly — git merge-tree --write-tree origin/main <branch> returns a
tree with 0 conflicts, and the local branch head equals origin's. The
correct action is awaiting_maintainer: nothing to do but merge it.

Acting on the emitted actions would demote an approved PR's issue and dispatch
a paid session to "resume" finished work.

Second, related defect: mergeable: UNKNOWN is treated as final

backlog-digest.sh:404 stores pr_state: ($pr.mergeable // null) from a
single query. GitHub computes mergeable lazily — the first query returns
UNKNOWN and triggers the computation. The sweep-prs skill documents this
explicitly and retries; the digest does not.

Measured on #490 during this pass: UNKNOWN on six consecutive queries spanning
~30 minutes, while git merge-tree proved it conflict-free the whole time.

Proposed fix

1. Link on any reference, not just closing keywords. Match #N anywhere in
   the PR body (Part of #N, tracking #N, a bare #N), and keep the closing
   keywords only for deciding whether the issue closes. Those are two
   different questions and only one of them is about linkage.
2. Retry mergeable until it leaves UNKNOWN, same as sweep-prs, and
   never let UNKNOWN masquerade as a definite state.

Acceptance

• With PR Consolidate PredictionSnapshotStore into the unified per-day DailyView format #490 open and unmerged, backlog-rhythm.sh emits no
  resume_implementation and no move_card for Consolidate PredictionSnapshotStore + ScheduleStore into the unified per-day DailyView format #409.
• An issue whose only PR says Part of #N derives In Review.