[S3] max(cs,1) floor silently inflates ratio metrics for zero-constraint windows

[erinepshovel-code]

constraint_statement_proxy_count returns max(cs, 1), so CS is always at least 1. This prevents a ZeroDivisionError (see retracted #45), but introduces a subtler problem: zero-constraint windows are silently treated as having one phantom constraint statement.

R_den = Rw / CS and L_den = Lw / CS are denominated against 1 instead of 0 whenever a transcript window contains no constraint markers. Metrics are emitted as if constraint presence was detected when it was not.

Impact: Short transcripts, informal dialogue, and early windows routinely produce zero constraint markers. Those windows' R_den/L_den values are not zero-signalled — they look like normal constraint-present windows with low Rw/Lw. Any downstream analysis that groups windows by constraint presence will conflate genuine low-constraint windows with constraint-absent windows.

Options:

1. Return a sentinel value (e.g. None, float("nan"), or a dedicated empty-window marker) when CS == 0 before flooring.
2. Add a companion boolean cs_present field to the window output so consumers can distinguish floored-zero from genuine constraint presence.
3. Document the floor explicitly in the module docstring so analysis consumers know CS is never 0 and adjust their interpretation accordingly.

Option 3 is the lowest-effort accurate fix; option 1 or 2 is the honest fix.