chore(e2e): add --user-id sticky routing + master test runner

[songkuan-zheng]

Summary

Two related improvements to the e2e harness, plus a missing tooling piece.

1. e2e/tools/call --user-id <id> — forwards to OpenAI user /
   Anthropic metadata.user_id, which the corp gateway uses for
   sticky upstream-key load balancing
2. All Anthropic cases now pass --user-id so cache-related
   assertions are deterministic
3. e2e/tools/run-all-cases — first-class master runner (was a
   /tmp scratch script); fixes a metric-snapshot bug; ships with
   per-case case_NN() functions for easy extension

Why --user-id matters

The corp Anthropic gateway (maasapi.corp.anispark.ai) round-robins
anonymous requests across multiple upstream API keys, each with its
own per-account cache namespace. Two byte-identical requests from
"no one" land on different upstream accounts; cache_read never hits.

We initially mis-diagnosed this as "gateway doesn't support cache
read". Prod evidence proves otherwise: spend_logs entries
e0ab6f96 and eec6a70d (same device_id=b0cb8e4d... in their
end_user blob) both hit cache_read=128k-135k tokens. Other
contemporaneous requests without device_id show cache_read=0.

Adding --user-id reproduces sticky routing in tests:

e2e/tools/call --provider anthropic --user-id user-42 \
    --cache ephemeral --ttl 5m --seed s1 --prompt-tokens 1500
# → first call: cache_creation=1827

e2e/tools/call --provider anthropic --user-id user-42 \
    --cache ephemeral --ttl 5m --seed s1 --prompt-tokens 1500
# → second call: cache_read=1827 ✓

Without --user-id the second call also returns cache_creation=1827
because the gateway routed it to a different upstream account.

Why the runner ships as a tool

The previous runner lived in /tmp — every reviewer who wanted to
verify the suite had to re-attach context to assemble the glue.
With it in-tree:

e2e/tools/run-all-cases               # full suite, ~$0.05 in providers
e2e/tools/run-all-cases --skip-paid   # only free cases for quick smoke

Exits 0 iff every case PASSes (SKIPs allowed).

Metric-snapshot bug in the old runner

The old snap() returned the last matching /metrics series. Once
later cases minted new virtual keys / teams, the "last matching line"
swapped from one series to another, and the before/after delta turned
nonsensical. Case 01 routinely false-failed.

New snap_sum() sums across all matching series; before/after stays
stable as the series set grows.

What changes

File	Change
e2e/tools/call	new --user-id flag → OpenAI user field
e2e/tools/run-all-cases	new tool — master runner with snap_sum, --skip-paid, case_NN() structure, pre-flight proxy check
e2e/cases/data/03_prometheus_anthropic_read.sh	new fixture (was inline in /tmp runner) — required GREEN, no longer SKIP-on-miss
e2e/cases/data/11_error_information_message_populated.sh	poll spend_logs up to 20s instead of flat sleep 2 (async logger lag fix)
e2e/cases/{01,02,03,04,08,09}_*.md	add --user-id to every Anthropic call
e2e/cases/README.md	document run-all-cases + new-case SOP
e2e/README.md	tools cheat-sheet adds run-all-cases + --user-id

Test plan

• e2e/tools/run-all-cases → 13/13 PASS (no flakes across 3 runs)
• Case 03 cache_read deterministically hits 1827 tokens with --user-id
• Case 01 no longer false-FAILs when runner is invoked after case 09
  has minted virtual keys/teams
• --skip-paid works: 05/06/13 marked SKIP, remaining 10 cases run
• Black 24.10.0 formatting clean on e2e/tools/call
• No code changes outside e2e/ — pure additive harness work

Out of scope

• Sending user_id from prod clients (would actually raise cache hit
  rate and lower bills — recommend tracking as a prod config change,
  not a harness change)
• Reconciling historical under-billed cache requests (separate work;
  PR feat(router): backfill missing cost fields from canonical entry for known models #4 fixes this going forward but doesn't backfill)