Bound evidence per kind instead of failing the packet

[gowtham0992]

Bounds collector evidence per kind so a large job log trims instead of
failing the capture.

Problem

EvidenceBuilder.validate_limits raised once the aggregate 400-line or
64 KiB ceiling was reached. That aborted build_packet_for_job, and
because collect_packets builds every packet in one comprehension, a
single oversized job discarded the packets for healthy jobs in the same
run.

Change

• Per-kind line and byte budgets, clamped by what the packet has already
  consumed, so the aggregate ceilings cannot be exceeded by construction.
• bounded_evidence_text keeps a deterministic prefix satisfying both.
• EvidenceBuilder.add returns a boolean instead of raising, so an item
  that sanitizes to nothing is skipped rather than killing the packet.
• Trimming sets sanitization.truncated and adds a missing_evidence
  declaration.

Budgets total exactly the aggregate ceilings: 3 log windows at 56 lines
and 12 KiB, one workflow snippet at 48 and 8 KiB, four diffs at 40 and
5 KiB.

Verification against run 31668261705 attempt 2

Packets built from live GitHub data, not fixtures.

	windows-smoke	package
evidence	8 items, 18,361 B, 284 lines	5 items, 9,684 B, 161 lines
packet	33,367 B	23,320 B
Bar validator	accepted	accepted

• Delivery IDs match Bar's server-side derivation for both.
• Byte-identical across two independent live collections.
• Diagnostic content survives: WinError 32, page-fts-v1.sqlite, and
  PermissionError in the Windows packet; Forced include not found
  and .linkignore in the package packet.
• Citation ranges stay true after trimming. A window covering original
  lines 100-159 trimmed to 56 lines cites 100-155, and the last retained
  line is exactly line 155.
• Budgets are per job. A 4,000-line log on one job does not shrink the
  other job's packet.

Full suite 1041 passed, ruff clean, actionlint clean, release hygiene
passed.

Known follow-ups, not addressed here

• sanitization.truncated is effectively always true, because per-kind
  caps trim routinely. Both packets above report truncation while using
  28% and 15% of the byte budget, so the signal does not distinguish
  a real ceiling from a routine per-item cap.
• validate_limits no longer guarantees at least one evidence item,
  while Bar rejects an empty evidence array.
• Job isolation is size-only. A build error or a delivery failure on one
  job still prevents the other job's packet from being sent.