fix(io): force UTF-8 on stdout/stderr and candidate writes

[diazMelgarejo]

Atomic PR 2 of 3 (PR 1, episodic-mirror, merged as #57).

Summary

Two small UTF-8 robustness fixes in .agent/tools/learn.py, both stdlib-only:

1. Force UTF-8 on sys.stdout/sys.stderr at import time (guarded by hasattr(..., "reconfigure") for older Python), so non-ASCII claim text prints correctly regardless of the host locale — matters most on Windows, where the default console codepage isn't UTF-8.
2. open(path, "w") → open(path, "w", encoding="utf-8") for the candidate JSON write, so it doesn't silently pick up the platform default encoding.

Rebased onto current master — the branch originally also carried PR #1's commits (stacked, per the original 3-PR plan); those are now dropped since #57 already merged that content, leaving only this PR's own diff.

Tests

Full suite passes on top of current master: 177 passed (+ 3 in .agent/tools/test_learn_episodic_mirror.py), 1 pre-existing unrelated failure (test_unknown_executable_is_a_structured_start_failure, a macOS PermissionError-vs-FileNotFoundError platform quirk unrelated to this change).

Note

Force UTF-8 encoding on stdout/stderr and candidate file writes in learn.py

• Reconfigures sys.stdout and sys.stderr to use UTF-8 with errors='replace' at module startup in learn.py, replacing unencodable characters instead of raising errors.
• The stage() function now explicitly passes encoding='utf-8' when writing candidate JSON files, replacing reliance on the platform default encoding.

^{Macroscope summarized 5e38f01.}

[diazMelgarejo]

Related, independent PRs from the same review pass — each stands on its own, no dependency between them:

• fix(recall): exclude superseded lessons from lessons.jsonl retrieval #60 — recall.py excludes superseded lessons from lessons.jsonl retrieval
• fix(io): use a context manager in _lesson_already_appended #62 — context manager in _lesson_already_appended (built on top of this PR's commit on the fork side, but targets master directly — the diff will collapse to just its own delta once this one merges)
• fix(upgrade): loop-* skill copy doubles the skills/ path segment #63 — unrelated bug found while reviewing this same area: harness_manager/upgrade.py's loop-skill installer doubles a path segment

cc @codejunkie99 — happy to answer questions on any of these, or squash/reorder however's easiest to review.