docs(lessons): add lessons 86 and 87 on blind test scaffolding

[dekobon]

Adds two lessons and one sub-example, from investigating #1091 and
#1054 — the two bug-labelled issues the last lessons pass left
uninvestigated.

Both turned out to be test-quality failures rather than the perf work
their titles suggest, and both describe ways a test suite can look
thorough while being structurally unable to see a defect.

Lesson 86 — a normalizing test helper blinds every caller at once

check_ops sorted operators_str and operands_str — the values
returned by the code under test — alongside the expectations, from the
helper's first commit. Ops::operators and Ops::operands were built
from HashMap keys and Rust's RandomState reseeds per map instance,
so the order differed between two parses in the same process.

The consequence reached users: bca ops printed byte-different output
for an unchanged input on consecutive runs. Eighteen per-language
tests parsed real source and compared real vocabularies, and not one
could observe it. The fix sorts in production and removes only the
actual-side sort, which turned those same eighteen callers into
ordering regression tests without writing one.

Lesson 87 — an assertion can be correct and still be about the wrong rows

last_emitted_metric_group_uses_closing_connector selected group lines
at three columns while metric groups sit six columns in. It matched
exactly one line — the metrics header — and its closing-connector
assertion was a true statement about that header. The test passed with
every metric group rendering a dangling |-, which is the precise
defect its name claims to prevent.

The same fix records why contains cannot work on this output at all:
a deeper rail ends with the shallower one, and sum / value recur
across groups, so a substring search accepts exactly the
mis-indentation under test.

Lesson 85 — third sub-example

cyclomatic.modified is the only nested metric object, so
dump_object's per-field truncate is the only thing keeping the fields
that follow a nested object on the right rail. Removing it mis-indents
cyclomatic's trailing sum / value in every dump — with all 2,869
lib tests still passing. An independent instance of lesson 85's thesis
from a different subsystem.

Verification

Reviewed before commit, which caught two false claims in the first
draft — both now fixed and re-verified against the tree:

• The caller count is eighteen, not the "~20" issue fix(ops): operator/operand order is nondeterministic across runs #1091 records.
  The draft had dropped the tilde and asserted it flatly, which is
  exactly the count-rot lesson 84 warns about.
• "without adding a test" was false — PR fix(output): make the dump walks linear and the ops listing deterministic #1093 added two tests in
  the same commit, and the 20-of-27 figure decomposes as eighteen
  callers plus those two.

Six further imprecisions were corrected: the ambiguous sort_unstable
referent, the unreproducible 27 denominator (now scoped to the tests
under ops and output::dump_ops as they stood then), an overstated
"deletes a property from the contract" (the order was documented as
arbitrary, so the helper accommodated a sanctioned non-guarantee), an
overstated equivalence with lesson 23, the undated 2,869, and a
truncate description that implied a dedicated line.

Every figure in both entries was checked against the repository rather
than transcribed from the issue bodies. Entries cite issues and PRs,
never commit hashes — this repository rebase-merges, so branch hashes
are orphaned on landing. rumdl check passes; the change is purely
additive (two new entries appended, one sub-example inserted), so no
existing lesson was renumbered and the number-based references in
.claude/skills/ and .claude/rules/ are unaffected.