fix(skills): the seven follow-ups from the #1226 review

[girishjeswani]

Follow-up to #1226, addressing the seven review comments there. One commit per item, a test per code change, run both as CI runs test files (python3 <file>) and under -m unittest discover.

• The four tests fix(skills): eight harness defects from a real-model eval pass #1226 added never ran in CI. They sat below the if __name__ guard, and CI runs each test file as a script. Both files now end with the guard and carry a test that reads its own source and fails if anything is defined below it again (An eval set over the storefront example, with verified goldens and a worked run #1210 hit the same thing). 113 -> 118 and 169 -> 176 tests as CI runs them.
• --rejudge --only overwrote the whole ledger, and --rejudge re-answered every case. Every gate reads a.rebuild; the bare flag left it false. imply_flags sets it, as --from already did, so a narrowed re-judge splices and the saved answers are reused. Help text updated.
• A rebuild after the spilled result file is gone scored zero entities. PERSISTED_STUB and OFFLOADED were two regexes and two readers for one note, disagreeing on the colon and the trailing period. saved_result() is the one helper: no note, a named file that cannot be read, or the body. result_text, offloaded_json and the transcript parser read through it; the missing-file case records no summary (unmeasured), not zero. The OSError handler no longer repeats the failed read.
• A --model-dir outside the repo pinned a dirty model as clean. git_sha now checks the status exit code (128 for a pathspec outside the repo) and pins nothing; a relative --model-dir resolves against --model-repo. The worked example in running-a-run.md passes --model-dir; the --model-repo help and the schema row no longer say the marker covers the whole repo.
• _norm treated a ; inside a string literal as a separator, so 'Books;Media' and 'Books Media' were one query. Only a ; outside quotes separates clauses now.
• build_run_package.py kept its own copy of the old result reader; it imports run_baseline.result_text.

Versions: @malloy-publisher/skills 0.1.25 and create-malloy-package 0.0.21 (skill content changed: the run reference and the ledger schema). Bundle regenerated.

[jswir]

Approving. The seven follow-ups are all addressed, and I ran the branch end to end on the storefront-tour set to check them:

• The three changed test files run clean the way CI runs them (python3 <file>): 186, 118 and 4 tests.
• A 12-case arm scored 10 of 12, with 95.5% recall. 6 of the 12 cases had a tool result saved to a file, so the merged reader was exercised for real.
• --rebuild reproduced the live run exactly, with no answerer spend.
• I moved one saved file aside and rebuilt: that call was recorded as unmeasured (rankedSummary: null), not zero, and the other cases' ledger lines were kept.
• --rejudge --only <qid> reused the saved answer byte for byte, ran only the judge, and left the other 11 cases' ledger lines identical.

Three things I found are inline. Each one changes a recorded result, so I'd fix them before merge if it's quick, or in a follow-up otherwise:

1. Semicolon fix: a query with an escaped quote, or an apostrophe in a comment, no longer matches its printed form. The harness then grades the wrong query. This is a regression against the old code.
2. Saved-to pattern: an ordinary result that mentions "written to " is replaced by that file's contents, and is then scored as zero entities.
3. --model-dir pointing at a folder that doesn't exist inside the repo still records a dirty model as clean.

Two smaller cleanups are inline too: the tautological assert, and the unused OFFLOADED alias with the leftover comment.

Two older problems, not from this PR and not blocking:

• The run viewer shows an unmeasured call as zero. build_run_package.py:375-382 turns rankedSummary: null into n_returned: 0 with an empty error. I checked it on the rebuilt run above: the unmeasured call shows in calls.csv as 0, the same as a search that found nothing, and empty_calls counts it. It now happens on every rebuild whose saved files are gone, which --rejudge now triggers too. An unmeasured column, or leaving n_returned empty, would fix it.
• A get_context call that errors is stored as zero results. One answerer left out scopes, and the MCP error was recorded as {resultCount: 0} with the error attached, rather than rankedSummary: null.