jsconsumer: retire comments the broker-semantics suite superseded

The coherence audit found the docs teaching things this branch and COR-1257's suite had already disproved. No behaviour changes except the one noted below.

TERM. Two comments still asserted that a bare Term does not settle a message. It does — measured across limits, workqueue and interest retention. The reason this package never Terms is that a Term settles with no RECORD of what was discarded, which is the surface the capture path exists to close. Stating it the old way invited someone to "fix" the settlement and find nothing wrong.

NAK-ERA COMMENTS. Four comments described dispositions their own assertions contradicted: the ladder path "naks with the ladder delay" while the test asserts nothing is disposed; a failed capture "Nak'd instead" while the test asserts zero Naks; and at the delivery cap "a dropped Nak is not a retry" when the point is that nothing redelivers past the cap at all. The retry path stopped disposing in fc30c0f; these were the last places still saying otherwise.

FLOOR+1. Removed the floor-2/blocker-6 figures a later session could not reproduce, and pointed at shouldQuarantine, which describes the mechanism instead: floor+1 names a message the consumer never receives when unmatched sequences sit below the blocker with nothing acked beneath them.

BLOCKER IDENTIFICATION. The package doc promised the breaker "dead-letters the message pinning a stalled floor" — the one thing it deliberately does not do, since consumer info does not carry the blocker's identity. Aligned with the disclaimer Settle already carries: a conjunction of independently true facts, which may quarantine a message that was failing alongside the blocker, bounded to one per stall window and replayable.

ACKWAIT VERSUS THE FIRST RUNG. The violation text taught the disproved model ("the first redelivery follows neither"). The server silently REWRITES the stored AckWait to BackOff[0] on the ordinary path, so the timing is knowable — it is the ladder. The check stays, for the two reasons that survive: the declared value never runs, and a pedantic-mode create rejects the mismatch rather than normalizing it, so a config this scaffold applies quietly fails to reconcile under NACK's controller.

RECOVERBY, deduplicated — the one non-doc change. NewRetry restated Schedule's RecoverBy-versus-delivery-budget comparison verbatim, wording included. It now asks Schedule and surfaces only the RecoverBy violations, since the Schedule it can build carries no ladder. Same rejections, one copy of the rule.

Also: the "run NewRetry once locally" advice pointed at the wrong function (the ladder checks live in Config.validate via Schedule, not in NewRetry); a schedule_test case still called a flat client ladder "the shape Track A actually ships", superseded when the server became the single ladder owner; a test comment named a function other than the one it sat on; and the rung-count violation printed MaxDeliver as a redelivery count, when MaxDeliver deliveries leave MaxDeliver-1 redeliveries. That last one had an assertion pinning the off-by-one, so the test moved with it.

README gains the measured-clock restart trade-off the package doc already carried: a per-process failure clock means unrelated restarts make the breaker fire late, never early.

Refs ENT-1601, ENT-1535, COR-1257.