Correct deep-review paper evidence semantics

[sauravsingla]

This PR applies the evidence-backed corrections from the final deep review without changing the frozen selector or experimental outputs.

Key corrections:

• fixes GraphBolt operation-fraction labels from 0.01%/0.1%/0.5% to the generator-defined 0.1%/1%/5%;
• corrects the historical false_full interpretation: the 33 labels are explicit-full/non-fallback cases, not wrong-arm or proven slower-full cases;
• records that all six 0.35 fallback events occur in the declared 220K cascade stress case and that the 72 real-graph observations have zero such fallbacks;
• adds retained-artifact cumulative latency results (42.7% reduction overall, 42.2% on real graphs) and guards them in the submission validator;
• replaces the dangling “exactness theorem” wording with an exactness contract and adds a compact BFS correctness invariant;
• makes the frozen selector mechanics more explicit in the manuscript;
• advances the immutable submission-freeze contract to pvldb-2027-submission-v3 so the corrected manuscript can be archived without moving v2.

No post-result retuning is performed. The existing author affiliation block is intentionally not guessed or fabricated; final institutional/independent affiliation remains a factual author-supplied submission field.