fix(e2e): validate Discord message fetches and bound rate-limit retries

[itismyfield]

Summary

Validate E2E Discord message-fetch responses and add bounded, read-only rate-limit retries. The Gate C run at 2026-09-05 09:42–09:43 KST became UNEVALUABLE: harness failure: a decoded upstream error object was returned under local HTTP 200 messages, then a caller iterated its string keys and failed on .get. One admission and one E-1 body were preserved; no automatic rerun occurred.

The production route's upstream-status handling is separate work in #5702. This patch changes no production Rust, configuration, deployment, scenario YAML or POST retry behavior.

Scope

Path	+	-
scripts/e2e/tui_relay/discord.py	107	24
scripts/e2e/tui_relay/test_assertions.py	2	0
scripts/e2e/tui_relay/test_discord_client.py	270	3
Total, 3 paths	379	27

Within 20 paths/+800. No cap exception. Valid raw/enveloped arrays and same-ID edits remain supported. Malformed responses report observed local HTTP status and a bounded body excerpt; wrapped retry_after is labelled inference, not a fabricated observed upstream status. Actual HTTP 429 can honor valid headers even with an empty/non-JSON body.

At most three GET attempts, with a 15-second retry budget and no shortened reset delay. This is not a strict total-transfer wall-clock cutoff: urllib provides socket timeouts rather than a total-read deadline. No scenario/run auto-retry is introduced.

Validation

• Original final-refetch failure: offline baseline RED → candidate GREEN.
• Actual429 header-only empty/non-JSON/invalid-UTF8 owner: intermediate RED → candidate GREEN.
• Client suite 23/23; existing CI assertion suite 78/78, including all 18 message-fetch owners; related offline suites 129/129. Counts overlap and are not summed.
• Scope/syntax/portable-path/hotfile checks pass; standalone bundle restoration and full fscks pass; candidate clean.
• Low-risk user policy: local Cargo/FULL omitted because src additions0 and no runtime-authority changes. Required CI remains mandatory before merge. No local live traffic.
• Head c81d7b0; tree 55465df4e30b50cdb208455e1f5f3404c12ca890; base6d2db2ee8f0835038c29f2fd44f4027fff95d60e.
• Implementation evidence SHA256 16d7b0d9944faa95d2944c97cc3b0f5cf076db7cc7ad1abb7ee607bf581467c0; full bundle SHA256 6945de79d4766088bed8820420f8f8276409fe21ccc7a02fb87998fb10a970a0.

Independent exact-head review

Both fresh reviewers bound the same full head/tree and returned VERDICT: CLEAN / PUSH_GO1:

• Astra/ultra r1: P0/P1/P2 = 0/0/0; review SHA256 68917e3f729256fa22fdfbf1b61ab8c28b625139f65f3dcd9681bb109d678db4.
• account1 Fable r1: P0/P1 = 0/0, six nonblocking P2 notes; review SHA256 0d291e8ebeff470774d1d0719ce498f6816fe3c9e02035c0000e2ff26bfa3e80.

Accepted compatibility disclosures: the first GET socket timeout is also capped at the smaller of the configured client timeout and the remaining 15-second budget. Empty 2xx bodies now fail explicitly rather than silently returning an empty list. CI imports the 18 new fetch owners, not all pre-existing client tests; existing send-prompt coverage gaps are not claimed closed. Numeric delay parsing retains Python float syntax. The intermediate header-only RED commit is preserved in a separate forensic bundle, not the final publication bundle.

Execution boundary

After merge, the user authorized exactly one supervised Gate C rerun from the new main harness source, subject to fresh preflight. Deployed leader/worker remain 6d2db2e, O100 only, channel1509350490461180105, no redeploy/config change. The supervisor polling delta is a separately reviewed local artifact, not part of this PR. A failed rerun is sealed and returned for user decision, not automatically repeated.

Refs #5071
Refs #5191
Refs #5702