[DO-NOT-MERGE] execution: hive eest ci to include all devnet shards

[taratorio]

Why

The previous glamsterdam-devnet shard filtered --sim.limit by the specific BAL-era EIP numbers (7708|7778|7843|7928|7954|7976|7981|8024|8037). That kept the shard fast, but it skipped every test in the eest_devnet fixtures that wasn't on that EIP list — including ~19k regenerated previous-fork tests (Paris/Shanghai/Cancun/Prague/Osaka) that exist precisely to verify those forks still pass under Amsterdam-era state changes. We were not exercising any of them in CI.

This PR closes that gap: full per-fork coverage of eest_devnet, every shard under the 30-min CI budget.

What

Replace the single EIP-filtered glamsterdam-devnet shard with 7 devnet shards mirroring the eest_stable per-fork split, plus two splits forced by what we found while bringing it up:

Shard	sim-limit	Tests	Failed	Duration
paris-devnet	.*/.*fork_Paris	2,744	0	11m 35s
shanghai-devnet	.*/.*fork_Shanghai	3,001	0	7m 52s
cancun-devnet	.*/.*fork_Cancun	12,362	0	17m 59s
prague-devnet	.*/.*fork_Prague	16,040	0	24m 03s
osaka-devnet	.*/.*fork_Osaka	16,750	0	26m 06s
amsterdam-a-l-devnet	.*tests/[a-l].*fork_Amsterdam	11,096	0*	23m 31s
amsterdam-m-z-devnet	.*tests/[m-z].*fork_Amsterdam	~10,291	est. 0	~25 min (est.)
total		72,284	1 (flake)	~2h 17m (cumulative)

Measured locally against bal@v7.2.0 / devnets/bal/7, shards run sequentially. In CI they run in parallel, so wall time is bounded by the longest shard (~26 min).

* amsterdam-a-l-devnet was verified separately with 0 failures, but the EIP-7928 test_bal_invalid_extraneous_entries flake (#21364) fired once during the combined fork_Amsterdam run that produced the totals row. The flake lives in tests/amsterdam/... so it falls in the [a-l] half — hence the max-failures: 1 budget there.

Two splits worth calling out

1. paris-devnet and shanghai-devnet are separate shards (instead of one paris+shanghai-devnet mirroring stable). The consume-engine plugin on devnets/bal/7 does not honour regex alternation in --sim.limit: empirically, ".*/.*fork_(Paris|Shanghai)" leaks ~5.8k fork_Cancun tests into the shard (verified by inspecting the testCases JSON). Single-fork patterns filter correctly. Worth chasing upstream, but splitting unblocks us today.

2. fork_Amsterdam is split into two shards by first-directory letter under tests/ ([a-l] vs [m-z]). The combined shard ran ~21k tests in ~52 min, over the 30-min target. The letter split gives a ~52/48 partition; both halves measured under 26 min locally. Character-class regex was verified locally to filter exactly (no leakage between halves).

Failure budget

amsterdam-a-l-devnet carries max-failures: 1 for the EIP-7928 test_bal_invalid_extraneous_entries flake (#21364) — the test lives in tests/amsterdam/... so it falls in the [a-l] half. Every other devnet shard is max-failures: 0.

Docs

.claude/skills/hive-test/SKILL.md updated to match the new shard list, max-failures table, and the alternation/letter-split rationale.

[taratorio]

let's wait for 1 Hive EEST job run before merging: https://github.com/erigontech/erigon/actions/runs/26396228185

[taratorio]

let's wait for 1 Hive EEST job run before merging: https://github.com/erigontech/erigon/actions/runs/26396228185

all green in 31mins, should be good to go

[taratorio]

let's wait for 1 Hive EEST job run before merging: https://github.com/erigontech/erigon/actions/runs/26396228185

all green in 31mins, should be good to go

2nd run had a flake, will investigate first before merging, let's hold off