DOC: Add the bind fast-path task note to devnotes

[fumitoh]

Files the working instruction that #269 was implemented from, alongside IOSpecLiteralTask.md and CoreRefactorDesign.md. Renamed from BIND-FASTPATH-PLAN.md to match the directory's CamelCase convention; the content is verbatim.

Docs only — no code changes.

Why keep it

It records what the implementation cannot:

• The measurements that motivated the change — 15,914,995 bind sites over the 28 lifelib-products models, and the finding that 99.999% of them enter at CellsImpl.call rather than Cells.__call__, which is why the fast path is in get_node() and not at the obvious call site.
• The decision table behind the guard, including the two results that are easy to get backwards: defaults do not need excluding, and POSITIONAL_ONLY does not need excluding.
• The two traps only adversarial review found — find_match() passing a list rather than a tuple, and ParamFunc.__slots__ = () shadowing the base tuple so _copy_other copied nothing. Neither is exercised by the 15.9M-call workload, so measurement alone would never have surfaced them.
• The approaches ruled out (§6), so they are not retried: lru_cache on _bind_args, caching bound keys per (args, kwargs), consulting obj.data before the key exists, and putting the guard inside _bind_args.

Where it differs from what landed

Two deliberate deviations, both recorded in 58f2aef's commit message rather than edited into the note:

• _bind_tails also fast-paths default-reliant short calls, which §3 marks as slow-path. That took the measured hit rate from the note's 99.22% to 100.00%.
• §5's ParamFunc question was resolved by fixing _copy_other to iterate Formula.__slots__, not by documenting the exception.

🤖 Generated with Claude Code