In short
RTL freeze is the point after which the design's logic changes only through a formal engineering change, so that physical design can start without being undone. It is a decision, made on evidence, and the hard part is not calling it but holding it.
What freezes, and what does not
At RTL freeze the register-transfer-level description of every block is placed under change control. Bugs will still be found; the freeze does not pretend otherwise. What changes is the cost of a fix: before freeze an engineer edits a file, after freeze a change goes through an ECO with an impact assessment, because physical design has started and a logic change can undo weeks of placement and timing work.
Verification does not freeze. Coverage closure continues against the frozen RTL, which is the point: the design stops moving so that verification can finish measuring it.
The evidence that says it is time
Freeze is often called on a date because the schedule said so. The honest criteria are about the design's state. Functional coverage should be close to its targets with a slope that will reach them in the weeks planned. Open bugs should be known and triaged, with none that would change an interface. Every block's interfaces should match the integration spec. The IP drops that the frozen RTL depends on should be the final ones, or the deltas known.
Freezing early to protect the physical-design start is tempting and usually expensive: each logic change after freeze becomes an ECO loop, and loops are what eat the sign-off weeks.
How a freeze thaws
Freezes rarely break in one event. They thaw: a small ECO, then another, then a feature that a customer asked for, each reasonable alone. After a dozen, physical design is chasing a moving target and the sign-off date has quietly moved by a month without anyone deciding it should.
Holding a freeze means every ECO is costed in days against the tape-out date before it is approved, and the running total is visible to the person who owns the date. An ECO that costs three days is a decision; twelve of them unseen are a slip.
Holding the freeze to the date
This is a commitment problem more than an engineering one. Tymeline holds the freeze as a commitment with an owner, reads the ECOs as they land, traces each one through physical design to the date it threatens, and puts the trade-off in front of the owner while it is still three days rather than thirty. The next program starts knowing which kinds of change broke this one's freeze.