Prune stale local pr-* refs left behind by closed, merged and superseded PRs

[dannymcc]

The harness checkout carries twelve local pr-* branches: pr-263, pr-280, pr-281, pr-288, pr-290, pr-291, pr-292, pr-294, pr-336, pr-337, pr-339, pr-346 (plus a mirrored harness/pr-* for each, at the same shas).

All but pr-346 correspond to PRs that are merged, closed or ruled superseded:

• Merged (squash): ci: bump actions/setup-python from 6 to 7 #263, deps: update psycopg2-binary requirement from >=2.9.10 to >=2.9.12 #280, deps: update coverage requirement from >=7.15.1 to >=7.15.4 #288, deps: update python-dotenv requirement from >=1.2.2 to >=1.2.3 #336, deps: update gunicorn requirement from >=26.0.0 to >=26.1.0 #337 — the five dependency bumps confirmed on dev under issue Clear the dependency-bump PR backlog as one batch after the 2026-09-02 release #353.
• Superseded: HU translation #339, Initial Hungarian translation #290 — Hungarian catalogue work long since overtaken by Still hardcoded or missing keys for translation #340/Hungarian catalogue mixes formal and informal register #350/Re-land the Hungarian translation additions from #339 against the current catalogue #352.
• Awaiting contributor at needs_work: Adds fuel level tracking to trips #281, Dev #291, May dev #292, MySQL/MariaDB + Auto cost per unit #294 — no desk-side work pending.

These cost real time. A stale pr-346 ref is what produced the ! [rejected] ... (non-fast-forward) fetch error that the desk misread as a merge conflict for several cycles. Stale refs also poison git branch --no-merged readings, which is already flagged in desk memory as a hint-not-evidence trap for exactly this reason.

Scope

1. Delete the local and harness/-mirrored refs for the eleven PRs listed above. Leave pr-346 alone — it is active this cycle.
2. Adopt force refspecs as the default for PR fetches: git fetch origin +pull/N/head:pr-N. The + makes a stale ref a non-event and removes the need for a pre-delete step entirely.

Acceptance criteria

• git branch --list 'pr-*' and git branch --list 'harness/pr-*' list only pr-346 / harness/pr-346.
• No branch belonging to an unmerged, still-open PR is deleted.
• The force-refspec convention is written down where the desk's PR-fetch steps live, so this does not recur.

Note for whoever picks this up: the planner shell is read-only git and cannot run git branch -D. This needs an officer with write access.

[dannymcc]

Checked the checkout: confirms your numbers — 12 local pr-* branches and 12 harness/pr-* mirrors at identical shas. Spot-checked the categorisation and it holds up:

• pr-263's tip commit isn't literally on dev (as expected for a squash merge), but dev carries the squashed equivalent (4753a43 ... (#263)) — merged, confirmed.
• pr-339 is a single HU-translation commit; dev already has it re-landed via ce0603f ... from #339 (#352) — superseded, confirmed.
• pr-346 has real, unmerged, in-progress work (CSV import duplicate-detection / price-per-unit calc) matching what's actively being worked this cycle — correctly left alone.

This is checkout housekeeping and a desk-process convention rather than a code change to may, so it's not something an engineer PR can land — no source or tests are touched. Passing it to the team lead to get an officer with write access to prune the 11 branches (re-verifying none map to a still-open PR immediately before deleting) and to fold the +pull/N/head:pr-N force-refspec convention into the desk's fetch process.