fix: pass encoding='utf-8' to Path.write_text on all text outputs

[Ori-Levi1]

Summary

Path.write_text() falls back to locale.getpreferredencoding(False) when the encoding argument is omitted. On Windows machines running a non-UTF-8 system locale (Hebrew/cp1255, Cyrillic/cp1251, Japanese/cp932, etc.), writing strings that contain characters outside the ANSI codepage raises UnicodeEncodeError.

This PR pins encoding='utf-8' on every write_text() call site that emits text content.

Reproduction (Windows + Hebrew locale)

pipx install evoskill
cd <empty-dir>
evoskill init

Crash:

File "...\src\cli\commands\init.py", line 355, in init_cmd
    (evoskill_dir / 'task.md').write_text(TASK_MD_TEMPLATE)
  ...
  File "C:\Python314\Lib\encodings\cp1255.py", line 19, in encode
    return codecs.charmap_encode(input,self.errors,encoding_table)[0]
UnicodeEncodeError: 'charmap' codec can't encode character '→' in position 141: character maps to <undefined>

TASK_MD_TEMPLATE contains → (U+2192), which is not representable in cp1255.

Files changed

File	Content written	Risk if not UTF-8
src/cli/commands/init.py:355	task.md (TASK_MD_TEMPLATE — contains →)	Crashes evoskill init immediately
src/cli/commands/init.py:368	state.json	Defensive — JSON values may include non-ASCII branch names
src/cli/report.py:62	run-<timestamp>.md (run report)	Reports often contain emoji/arrows from agent traces
src/harness/opencode/options.py:107	OpenCode JSON config	Defensive — paths/instructions may include non-ASCII
src/harness/opencode/skill_utils.py:76	Evolved SKILL.md (frontmatter + body)	Will crash on any evolved skill containing arrows, em-dashes, smart quotes
src/loop/helpers.py:227	Evolved prompt files	Will crash on any prompt the proposer generates with non-ASCII
src/loop/runner.py:196	loop_checkpoint.json	Defensive — checkpoint values are numeric but should be UTF-8 explicit

Each change is write_text(x) → write_text(x, encoding='utf-8'). No behavior change on UTF-8-default systems (most Linux/macOS); fixes the crash on Windows non-UTF-8 locales.

Why this matters beyond the immediate crash

Even if a user gets past evoskill init (e.g., on a US-English Windows install), the SKILL.md and prompt files are written by the LLM-driven proposer. The proposer routinely emits Unicode (em-dashes, arrows, curly quotes) when it rewrites a skill. So evoskill run is also affected on any non-UTF-8 locale, just on a less predictable iteration. Pinning UTF-8 once removes the entire class of bugs.

Out of scope

Two more write_text() call sites exist that were not patched here because I didn't exercise their code paths during reproduction:

• src/docker/launcher.py:111 (compose file)
• src/remote/base.py:28 (remote run record)

The same change is recommended in a follow-up.

Test plan

• Manual: evoskill init succeeds on Windows 11 + Hebrew locale (cp1255) + Python 3.14
• Manual: evoskill run writes SKILL.md with em-dashes and arrows without error
• No new tests added — change is a one-character explicit-encoding hint per call site, behavior identical on UTF-8-default systems