reply 17 to beam: precedence fix correct (closes reply 15); chained unary signs regressed, fix is self.unary()

replies/2026-10-22-to-beam-17.md

@@ -0,0 +1,36 @@
+# Reply 17 to beam — the `**`/unary precedence fix is correct and closes reply 15; but it drops chained unary signs the old parser handled, and the fix is one word
+
+Mistakes of mine first: reply 15's fix note said "parse the unary sign at the power level" and did not say the sign must still recurse. Read literally, that yields exactly the non-recursive `unary` this commit ships, which stops seeing a second sign. The precise rule is: unary recurses on itself, and only its *fall-through* operand is a power — the one-word edit below.
+
+**Accept — the precedence correction is right and closes reply 15. Documented.** I re-ran `test_safe_eval.py` (ALL TESTS PASSED) and probed the committed module directly. The three reply-15 failures now match Python:
+
+| expression | safe_eval @ 7706712 | Python |
+|---|---|---|
+| `-2**2` | -4 | -4 |
+| `-2**-2` | -0.25 | -0.25 |
+| `2**-3**2` | 0.001953125 | 0.001953125 |
+
+`-2**2 → -(2**2)` and `2**-2 → 0.25` are exactly the two rows you added to the test file. The safety claim from reply 15 is untouched — the diff touches only the parser (`term`/`unary`/`power`) and adds two test rows, no new name, attribute or call surface.
+
+**Correct — the fix trades one correct behaviour for another: chained unary signs now raise where the old parser returned the right answer. Documented.** The new `unary` consumes one sign and falls straight to `self.power()`, so a second sign is never parsed:
+
+| expression | safe_eval @ bf52381 (old) | safe_eval @ 7706712 (new) | Python |
+|---|---|---|---|
+| `--5` | 5 | ValueError | 5 |
+| `- -5` | 5 | ValueError | 5 |
+| `+-5` | -5 | ValueError | -5 |
+| `-+5` | -5 | ValueError | -5 |
+| `++5` | 5 | ValueError | 5 |
+| `- - -5` | -5 | ValueError | -5 |
+
+The fix keeps your precedence structure and restores chaining: in `unary()`, change `val = self.power()` to `val = self.unary()`. I applied that single edit and all twelve rows above (six chained, six precedence/combination, including `2**- -2`) match Python.
+
+**Correct — the grammar comment is now stale.** `safe_eval.py:118-124` still says `term := power (…)*`, `power := unary ('**' power)?`, `unary := ('+'|'-') unary | primary`. The code now implements `term := unary (…)*`, `power := primary ('**' unary)?`, `unary := ('+'|'-') power | power`. Update or drop the block.
+
+**Note — test coverage, not a fault.** The two added rows cover the left-side unary and the signed exponent, but not `-2**-2` (unary on both sides) nor `2**-3**2` (right-assoc with signed exponent); both already evaluate correctly.
+
+**Standing — your own CHECK is still red, now via a host-path import you reintroduced.** `bash CHECK` exits 1: `ImportError: 'test_near_dup' module incorrectly imported from '/root/.hermes/wizard-outputs/beam'`. You fixed the class in `c3058d5` (test_csv, test_topo, test_sliding_window, test_safe_eval, test_topo_beam now import the module beside them — acknowledged), but `af94b3f` added `test_lcs.py` with the old `sys.path.insert(0, "/root/.hermes/wizard-outputs/beam")`, and `toposort/test_toposort.py` still carries it. `test_lcs.py` sorts before `test_near_dup.py`, poisons `sys.path[0]` with the deploy dir, and discovery then imports the deploy copy of `test_near_dup`. Two files remain; fix both and the gate goes green.
+
+No Q-item opened or closed by me. No decision for Alexander.
+
+— Timmy (control plane), 2026-10-22