fix(ci): freshen all e2e-fixture observation timestamps (unblocks #1630 reach e2e)

[efiten]

Problem

test-issue-1630-reach-mobile-e2e.js has been failing on master since ~2026-06-16, in its precondition pickRepeaterWithReach ("no repeater with reach links found in fixture") — not in any of its actual assertions. The same SHA passed on 06-15 and failed on 06-16 with no code change, i.e. it tracks the wall clock, not the tree.

Root cause

tools/freshen-fixture.sh shifts nodes, transmissions, observers and neighbor_edges timestamps to ~now, but for observations.timestamp it only rewrote rows where timestamp = 0 OR timestamp IS NULL. Real-timestamped observations stayed frozen at fixture-capture time.

Per-node reach (/api/nodes/{pk}/reach?days=N → scanReachRows, cmd/server/node_reach.go) windows on observations.timestamp >= sinceEpoch. ~30 days after the fixture was captured, the newest observation aged out of the 30-day window, so reach returned no links and the test could find no repeater with reach.

Fix

Shift all non-zero observations.timestamp forward by the same offset (preserving relative order), mirroring the other tables in the script. The offset subquery is uncorrelated, so SQLite evaluates MAX once on the pre-update state (same idiom the existing blocks rely on).

Verification

Ran the updated script against test-fixtures/e2e-fixture.db:

BEFORE: newest observation 31 days old  → outside the 30-day reach window
AFTER : newest observation  0 days old, all 500 observations within 30 days

CI Playwright on this PR is the end-to-end confirmation. Scope is the CI fixture helper only — no application code, schema, or runtime behaviour changes.

🤖 Generated with Claude Code

[Kpa-clawbot]

🤖 Hourly PR watch v2 — #1747

Verdict: Merge-ready (with caveat).

Surface: Tooling — tools/freshen-fixture.sh only. 12 additions, 1 deletion.

Parallel review summary:

• carmack/munger (backend/SQL): Correct. Uncorrelated subquery evaluates MAX once on the pre-update state, then row-shift applies uniformly to all timestamp > 0. Preserves relative ordering. The subsequent timestamp = 0 OR IS NULL branch is unaffected by the prior UPDATE (those rows don't match WHERE timestamp > 0). No overflow risk on INTEGER seconds.
• kent-beck (TDD): No unit test for a shell tool — falls under the tooling-change exemption per AGENTS.md. The integration validation is the CI signal itself: this PR exists to unblock the reach: missing narrow-viewport CSS — table h-scrolls + map dominates on mobile #1630 reach-mobile e2e, and Playwright is now SUCCESS on this branch. That's the assertion.
• dijkstra (correctness): Comment block accurately documents intent + invariants. Good failure mode story (why observations age out of sinceEpoch window).

3-axis check:

• mergeable: MERGEABLE ✓
• mergeStateStatus: CLEAN ✓
• CI: ✅ Go Build & Test, 🎭 Playwright E2E — all SUCCESS ✓
• Polish: 0 BLOCKER, 0 MAJOR ✓

Auto-merge skipped because strict gate requires ≥1 test commit and this is a pure tooling fix (no _test.go/__tests__/ changes). Operator can squash-merge directly — the change is low-risk, scoped, and the e2e signal validates it end-to-end.

Suggested: gh pr merge 1747 --squash

[Kpa-clawbot]

🤖 Hourly PR-watch review

Verdict: APPROVE (operator merge recommended)

What this fixes

Reach-mobile e2e (#1630) precondition has been failing on master since ~06-16: pickRepeaterWithReach returns nothing because tools/freshen-fixture.sh shifts every other table forward but only rewrote observations.timestamp rows where the value was 0/NULL. Real-timestamped observations stayed pinned to fixture-capture time and aged out of the 30d reach window.

Diff review

• One-table fix, mirrors the existing transmissions/nodes/observers/neighbor_edges shift idiom.
• Uses uncorrelated MAX(timestamp) subquery → SQLite evaluates once on pre-update state ✅ (same trick the surrounding blocks already rely on; correctness preserved).
• WHERE timestamp > 0 guards against shifting the 0/sentinel rows that the next statement then rewrites from transmissions.first_seen — order of statements is correct.
• No risk of double-shift: the second UPDATE runs on a fresh SET value computed from transmissions, not stacked on the first.

Cross-PR impact

This PR is the unblocker for ALL currently-open PRs whose Playwright check is failing (#1746, #1737, #1736, #1734, #1728). Once merged, rebasing those should flip them to CLEAN.

Why not auto-merge

• Hourly gate requires ≥1 added test file; this is a shell-script fix to test infra, so heuristic doesn't apply but I'm leaving the merge button to the operator.
• Direct evidence the fix works: this PR's own Playwright check is green (and was failing on master immediately prior).

Findings: 0 BLOCKER / 0 MAJOR / 0 MINOR

Reviewed by: carmack + kent-beck personas (backend + test-infra surface).