fix: parse integral scientific notation in CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS

[azmi2409]

Fixes #1359.

run_end_ceiling_ms() parsed CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS with a
bare int(raw), so CLI-compatible scientific-notation spellings like "1e6"
were treated as invalid and silently fell back to the 600_000ms default
instead of 1_000_000.

The value is now parsed by a small _parse_run_end_ceiling_ms helper that
accepts plain integers (with surrounding whitespace, as before) plus integral
scientific-notation spellings such as 1e6, 1E6, 2.5e3. Everything else
keeps the old behavior:

• unset / invalid strings / empty -> 600_000 default
• "0" -> 0 (no limit)
• negatives -> default
• non-integral values (1.5, 1e-1) -> default
• very large values are honored, then clamped downstream by the existing
  _MAX_RUN_END_CEILING_MS sleeper bound

The parser also guards against absurd exponents (1e999999999) so it can
never build a gigantic int.

Adds a "1e6" -> 1_000_000 regression case to
TestRunEndCeilingFromEnv.test_parse. Full suite: 1588 passed, 6 skipped;
ruff check, ruff format --check, and strict mypy are clean.

[sigley]

I found one boundary case in the exponent guard on the current PR head (924746e).

Running the parser from this head on Python 3.10.11 gives:

_parse_run_end_ceiling_ms("0e18") -> 0
_parse_run_end_ceiling_ms("0e19") -> 2147483647
run_end_ceiling_ms({"CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS": "0e19"}) -> 2147483647

That changes zero from the documented 0 = no limit semantics into the maximum finite sleeper ceiling solely when the exponent crosses 18. The cause looks like the shift > 18 fast path returning _MAX_RUN_END_CEILING_MS before the zero coefficient is handled.

Would it make sense to handle coefficient == 0 before that large-shift guard and add a 0e19 -> 0 regression case? That should preserve the zero semantics while keeping the bounded path for non-zero absurd exponents.

[azmi2409]

Good catch — fixed. The zero coefficient is now handled before the large-shift guard, so 0e19 (and any 0eN) keeps the documented 0 = no-limit semantics. Added a 0e19 -> 0 regression case (zero_scientific_notation_boundary) to TestRunEndCeilingFromEnv.test_parse. All 10 tests in the class pass.