workflow-fix: UnicodeDecodeError escapes 112 json-guard except tuples (53 in the watcher, all on read_text paths)

Overview / Motivation

except (json.JSONDecodeError, OSError) wrapped around a Path.read_text() + json.loads() pair does not catch UnicodeDecodeError, which read_text() raises on any file that is not valid UTF-8. UnicodeDecodeError subclasses ValueError, not OSError, and json.JSONDecodeError only fires after decoding has already succeeded — so a single encoding-corrupt JSON file propagates an uncaught exception out of a guard that was clearly written to be total.

Surfaced during #2164 round 2. That task fixed exactly one instance (autonomous_session_watch._stalled_cap_gpu_hours) because its own change had newly routed that helper onto the watcher's unwrapped gate-push pass, where a corrupt ~/.eps-autonomous/issue-<N>.json would kill an entire watcher tick. The #2164 implementer flagged one sibling. The #2164 orchestrator then measured the actual scope before filing.

Measured scope (2026-08-07, origin/main 68fbf9bf3e)

• 53 occurrences of the two-element tuple in scripts/autonomous_session_watch.py.
• 53 of 53 have a read_text() call within 8 lines above the handler — i.e. every single one is on a decode path and carries the hole.
• 112 occurrences repo-wide across scripts/ + src/ (*.py).

Reproduce:

grep -c "except (json.JSONDecodeError, OSError)" scripts/autonomous_session_watch.py   # 53
grep -rn "except (json.JSONDecodeError, OSError)" scripts/ src/ --include="*.py" | wc -l  # 112

Goal

Close the UnicodeDecodeError escape everywhere the pattern guards a read_text() + json.loads() pair, so a corrupt-encoding state file degrades to the intended fallback instead of propagating.

Proposed change

Add UnicodeDecodeError explicitly to the tuple at each decode-path site: except (json.JSONDecodeError, OSError, UnicodeDecodeError). #2164 set the precedent at _stalled_cap_gpu_hours and deliberately chose the narrow, explicit form over broadening to bare ValueError — keep that, so the handler still cannot swallow an unrelated ValueError that should propagate.

Mostly mechanical, but not blindly sed-able. Two judgement calls:

1. Severity triage. A site whose caller is already wrapped is latent; a site on an unwrapped pass (the #2164 case) is live. Prioritize and state which is which rather than treating all 112 as equal.
2. Is fallback right here? At a few sites, silently falling back on a corrupt file may be the wrong call and a loud failure is better — "fail fast, never hide failures" applies. Where the guard's existing OSError/JSONDecodeError branch already returns a default, matching that for UnicodeDecodeError is consistent; where it re-raises or logs, match that instead. Do not convert a loud path into a silent one.

Acceptance criteria

• Every decode-path site in scripts/autonomous_session_watch.py (53) handles UnicodeDecodeError consistently with its existing fallback behavior.
• The wider scripts/ + src/ set (112 total) is either fixed in the same sweep or explicitly triaged with a recorded reason for anything deferred — a partial sweep must say what it left and why, not go quiet.
• At least one regression test per distinct guard shape (not per site): write an invalid-UTF-8 file (b"\xff\xfe..."), assert the fallback is returned rather than an exception propagating. #2164 shipped exactly this shape as test_reported_cap_falls_back_on_encoding_corrupt_registry_entry — reuse it as the template.
• No site is converted from loud-failure to silent-fallback as a side effect.
• A lint or test guard preventing the two-element form from being reintroduced on a read_text() path would be worth more than the sweep itself. Consider it as the primary deliverable; #2164's test_cap_env_read_is_single_sourced is the in-repo precedent for a source-scan invariant test.

Notes

json.JSONDecodeError is itself a ValueError subclass, so any site that already catches bare ValueError is fine and should be left alone. Check before editing; the grep counts above are the raw two-tuple form only.