EIP-9999: amend EIP-8141 mode check, fix discard + mode-1 binding, spec-consistency fixes

[soispoke]

This PR makes a set of small, text-provable spec fixes to eip-9999.md. It targets the patch-3 branch so the fixes flow into ethereum#11772. Design choices (dependency canonicalization, the wrapper cap model, proof placement, gas accounting, FOCIL binding) are not changed here; they are listed at the end as discussion comments.

Correctness

1. EIP-8141 mode check. EIP-8141 enforces assert frame.mode < 3, so DEP_VERIFY_FRAME_MODE = 3 frames are rejected and the mechanism cannot validate as written. Amended to < 4. (The one fix without which nothing validates.)
2. Discard set was dead code. The recursive circuit built discard_hashes as a set of hashes but tested raw dependency triples against it, so no discard ever applied (Test Case 10 could not pass). Now compares triples to a set of triples.
3. Mode-1 wrapper binding. Mode-0 wrappers verify deps equals the referenced transactions' dependencies; mode-1 did not, so a mode-1 wrapper could carry a valid recursive STARK over deps unrelated to its transactions. Added the same binding to mode 1.

Clarifications

4. Tx-by-hash wrappers require the node to hold the full transaction bytes before the wrapper can be validated or forwarded (needed to recompute dependencies(tx)).
5. No-fold forwarding fallback. A node that cannot produce a fresh recursive wrapper before the next interval SHOULD forward the wrappers it already verified, so folding stays an optimization rather than a liveness gate.
6. AGGREGATED_VK propagation. Stated that AGGREGATED_VK is a protocol constant, never user-chosen, pinned by block validity rule 2 and carried unchanged through every recursion level (which is why a recursive proof's public inputs are [deps_hash, AGGREGATED_VK] while a leaf takes only data_hash); distinguished it from the per-dependency user vks in the Security note.

Editorial

7. "two new frame modes" → "one new frame mode" (only one is defined).
8. FRAMEDATASIZE is not an EIP-8141 instruction; use FRAMEPARAM(0x04, i) for length (matches this EIP's own introspection list).
9. Unified get_dependencies(block) / dependencies(block) to the defined name.

---

Design comments (out of scope for this PR — flagging for discussion)

These are real but change semantics or the DoS model, so they are not in this bug-fix PR:

1. Dependency canonicalization. get_deps_hash(filtered_deps) == block_deps_hash cannot hold when a block has a duplicate (scheme, data_hash, vk) triple (the block hash counts it twice, the circuit dedups to one) or when the proof is assembled by recursive union (deps ++ inner_deps order ≠ block-tx order). Likely fix: make dependencies a canonical set (sort by the 96-byte dep_id and dedup) in both block_deps_hash and the circuit. This changes consensus semantics, so flagging rather than patching.
2. Mode-0 vs mode-1 cap axis. MAX_LEANSTARK_DEPS_PER_WRAPPER = 1 (a Mempool-Level Limit) means a mode-1 mempool wrapper can't aggregate more than one STARK-bearing tx, contradicting "one wrapper containing all currently active transactions". A recursive proof is ~constant size, so leaf count is the wrong axis. Suggest: per-dep caps apply to mode-0 raw wrappers; mode-1 wrappers bounded by serialized size + a tx-reference count.
3. Per-tx caps not enforced. MAX_SIGS_PER_TX / MAX_STARKS_PER_TX are described in Security as the primary DoS defense, but no validity rule enforces them, and MAX_DEPENDENCIES_PER_FRAME = 256 lets one frame exceed both.
4. Gas. recursive_stark_gas appears to charge each dependency twice (per-frame + block-level) and bills users via base fee for the builder's work — clarify additive-vs-replacing and the payer. Separately, make the proof-size bound explicit (fixed trace width, 2**24 cycle cap) so flat gas is visibly justified. (Not proposing dynamic gas.)
5. Proof placement. The header carries the full proof inline; consider committing only to a recursive_stark root + block_deps_hash and carrying the proof bytes in a separate erasure-coded object (ePBS commits the small header first, sends the body later).
6. FOCIL. FOCIL validation doesn't state attesters must check deps_hash == hash(deps of the FOCIL's txs) and verify against AGGREGATED_VK (the same hole Spelling, Grammar, Initial Structure ethereum/EIPs#3 closes for mempool wrappers); and FOCIL = [transactions, recursive_stark] intentionally exceeds EIP-7805's 8 KiB IL budget — worth one line documenting that accepted tradeoff.
7. Rationale notes: mention mempool sharding as future work; note that aggregation incentives are a deferred open problem; and reconcile "unlimited-depth recursion" with the depth-2 soundness assumption.