The exemplar replay ignores order, indent, and four of five console blocks

[CameronBrooks11]

What

test_the_spacer_exemplars_diff_transcript_is_what_the_console_prints
(tests/test_docs.py) now requires every quoted output line to be a line the
tool printed, and #376 added a floor so a step cannot quote nothing. Three
weaknesses remain, all found by adversarial review of #376 and all confirmed
passing on both pinned engines:

1. Order and multiplicity are not asserted. The check is set membership,
   so swapping the two quoted lines, or claiming one printed twice, passes.
2. Leading whitespace is discarded. Both sides are .strip()ed, so
   un-indenting the covered: continuation line passes — and that indent is
   how the README shows it is a continuation.
3. Only the first partspec diff console block is replayed. The
   exemplar's other four blocks — the --pin, the PASS: 8 pass / echo $?
   pair, the exit-4 pin mismatch, and the --pin rewrote claims stderr — are
   asserted by nothing. Falsifying one line in each leaves the whole file
   green, and just example-spacer does not cover them either.

Why it matters

#361 existed because the exemplar README drifted from what the tool prints and
three gates stayed green over it. The replay closes that for one command. The
other four commands in the same file are still in the state #361 was filed
about.

Smallest fix

For (1) and (2): compare the quoted lines to the tail of
printed.splitlines() in order, without stripping the left side.
For (3): replay every console block, not next() of them.

Found by review of #376, deliberately left out of that PR's scope as wider
than #361.

[CameronBrooks11]

(1) and (2) done in #393; (3) is bigger than "replay every console block"

Parts (1) and (2) are fixed: the replay now matches quoted lines as an ordered
subsequence of produced lines with the indent intact. Mutation-proved — swapping the two
lines, claiming one printed twice, and un-indenting the covered: continuation all fail
now and all passed before. Elisions stay allowed, since the gaps between matches are free.

Part (3) I have not done, because the smallest fix as written does not work. Looked at
the file rather than the count:

There are seven console blocks, not five, and blocks 1–6 are one chained narrative
against shared state, not six independent runs. Replaying them needs three things the
replay has no notion of:

1. A block with no command at all. The exit-4 pin-mismatch block opens with output.
   Its command lives in the prose above it: "Now loosen a bound — PLATE = (40.0, 30.0, 9.0) — and re-run the same command". So the replay would need to read a prose-encoded
   edit and a prose-encoded "re-run the previous command" — two forms it does not
   parse, and neither is a $ line.

2. A literal elision marker. That block quotes ... elided ..., which is not a line
   the tool prints. The current single-block replay never meets one. Any all-blocks replay
   has to distinguish an elision marker from a stale line, and getting that wrong turns
   this gate into the thing it is guarding against.

3. State carried between blocks. Block 5's --pin rewrote claims warning only appears
   because block 4's edit happened; block 6 reads the report.json block 5 wrote. Replayed
   independently, three of the six assert nothing or fail for the wrong reason.

So next() is not the whole distance between here and correct. The options, none obviously
right:

• Teach the replay the prose forms — an edit directive and "re-run the same command",
  plus an elision marker. Most faithful to the README as written, and it means the gate
  depends on parsing English in a document whose whole problem is that English drifts.
• Restructure the exemplar so every block is self-contained and every command is a $
  line. Makes the replay trivial and changes a document written for humans to suit a test.
• Replay blocks 1–6 as one chained session with the edit expressed as a # directive
  the way block 7 already does (# ... now edit BORE_D: 8.0 -> 12.0 ...). This is the
  smallest honest version: block 7 already proves the form works, and it would need one
  line of README changed rather than a restructure.

The third looks best to me, but it edits the exemplar to suit the gate, which is a real
trade and not mine to make silently. Leaving (3) open with this scoping rather than
shipping a partial replay that looks like coverage.

What is still ungated after #393: four commands in this file — the --pin, the
PASS: 8 pass / echo $? pair, the exit-4 pin mismatch, and the --pin rewrote claims
stderr. Exactly what this issue said, unchanged.