feat(morpho-v2): group per-vault alerts into one Telegram message; sync KATANA V2 vaults

[spalen0]

Summary

Two changes bundled because they unblock the same workflow.

A KATANA Gauntlet USDT increaseTimelock(setSendAssetsGate) landed without a follow-up "executed" alert because the vault wasn't in VAULTS_V2_BY_CHAIN for KATANA. And even for vaults we do track, a 22-call batch operation (e.g. the same InfiniFi LongTimelock upgrade that surfaced the Wavey Gist bug) would have spammed 22 separate "pending config" Telegram messages.

1. Add KATANA V2 vaults the team already tracks in prod
   protocols/morpho/_shared.py: Gauntlet USDT, Steakhouse High Yield USDC, Steakhouse Prime USDC on KATANA. Addresses and risk levels mirror /srv/monitoring production config — the only place these were already tracked. Result: they get the same pending/executed/role/set-diff treatment as mainnet+BASE.

   Note these three are curated by Gauntlet and Steakhouse, not Yearn, so the "Yearn-curated" wording on VAULTS_V2_BY_CHAIN (and the matching docstrings in markets_v2.py) is corrected. The list is shared with markets_v2.py, so adding a row here also enrols the vault in market/allocation checks — that's intended, and now stated in the comment.

2. Group per-vault alerts into one Telegram message
   protocols/morpho/governance_v2.py: every diff category (pending new, pending resolved, owner, curator, sentinels, allocators, adapters) now buffers into a per-vault _VaultDiff, then diff_and_alert flushes ONE Telegram message per vault with:

   • one header (V2 [name](url) on chain)
   • one severity (highest of the group: CRITICAL > HIGH > MEDIUM > LOW)
   • sections separated by ---

   Worst case (InfiniFi 22-call batch) goes from a 22-message Telegram burst to 2 numbered parts (see below).

Message length

Telegram caps a message at 4096 chars, and utils.telegram truncates past that (and drops Markdown with it), so naively concatenating would have lost the tail of exactly the batch this PR exists to fix — the 22-call case renders ~6.3k chars, i.e. 7 of 22 sections silently dropped.

_send_vault_alerts therefore packs sections into as many messages as needed, each under budget, with the header repeated and an (i/N) suffix:

messages: 2   lengths: [3783, 2649]   sections delivered: 22 / 22

Delivery vs. cache

Cache cursors are buffered in _VaultDiff and committed only after the send succeeds. Writing them during the diff pass would mark a change as alerted even when Telegram failed — main() logs the exception and moves on, so the alert would be lost permanently. A partial send (part 1 lands, part 2 fails) now repeats the whole group next run; duplicates beat a governance change nobody ever sees.

Before / after

Before (each diff category fires its own Telegram message):

⚠️ ⏳ V2 Gauntlet USDT on katana
📥 Submitted: increaseTimelock(setSendAssetsGate → 604800s)
⏰ Executable at: 2026-08-06 19:11:34 (6 days)
🔗 Tx: 0x603be77d…

ℹ️ ✅ V2 Steakhouse Prime USDC on katana
Pending operation IncreaseAbsoluteCap() was executed (was due 2026-07-30 20:32:41).

ℹ️ ✅ V2 Steakhouse Prime USDC on katana
Pending operation IncreaseAbsoluteCap() was executed (was due 2026-07-30 20:32:41).

After (one grouped message per vault):

🚨 V2 [Gauntlet USDT](https://app.morpho.org/katana/vault/0xaC59…) on katana

---

📥 Submitted: increaseTimelock(setSendAssetsGate → 604800s)
⏰ Executable at: 2026-08-06 19:11:34 (6 days)
🔗 Tx: [0x603be77d…](https://katanascan.com/tx/0x603be77d…)

---

🚨 👑 Owner changed: `0xold` → `0xnew`

---

🧩 adapters changed
  + `0xdd…`
  − `0xee…`

Tests

• Existing test_morpho_v2_governance.py updated for the buffer-based API (_diff_* / _alert_* now take a _VaultDiff); the pending-label test drives the public diff_and_alert instead of the private helpers.
• New test_multiple_alerts_for_one_vault_are_grouped_into_single_message: 3 pending + 1 owner + 3 set changes produce exactly 1 alert at HIGH severity, with the vault header once and all 3 pending Txs present.
• New test_no_alerts_emits_no_message: vaults with no diffs stay silent (no spam).
• New test_only_low_severity_changes_emit_low_severity_alert: pure allocator/pending-resolved diffs are LOW, not the HIGH of the previous "first-message-wins" style.
• New test_oversized_group_is_split_into_numbered_parts_without_dropping_sections: the 22-call batch splits into >1 message, each ≤4096 chars, with all 22 sections and tx hashes present.
• New test_cache_writes_are_deferred_until_the_send_succeeds: a raising send_alert leaves the cache untouched; a working one commits the cursors.

All 657 tests pass; ruff format and ruff check clean.

Note

Based on feat/peg-monitoring-l3-events (#307) — retarget to main once that lands.