perf(output): remove the dump walk's quadratic prefix and parent lookups

[dekobon]

Fixes #1054

bca dump was quadratic in AST nesting depth. The issue is already
closed with the full write-up; this PR is the code.

Root cause: two quadratic terms, not one

1. The prefix, as the issue describes. Every queued node carried an
   owned String copy of its ancestors' box-drawing prefix, rebuilt with
   format! per node and clone()d per child. A tree that is wide and
   deep holds O(depth²) resident bytes plus an O(depth) copy per node.

2. Node::parent, which the issue does not mention. branch_glyphs
   called node.parent().is_none() once per node for the flush-left
   check, and tree_sitter resolves parent() by descending from the
   root — the same shape perf(metrics/tokens): O(leaves x depth^2) ancestor walk is a DoS vector #1052 and perf(metrics): Node::parent lookups make three walks quadratic in nesting depth #1084 hit. This term dominates on the
   wide-and-deep fixture.

What changed

The walk keeps one shared prefix buffer, extended on descent and
truncated at the start of the next visit; each stack frame carries a
prefix_len instead of a String. Truncating on visit rather than on
the way back up is what makes a bare length sufficient — prefixes only
grow as the walk descends, so the first prefix_len bytes are still
this node's prefix when it is popped.

The flush-left decision moved onto the stack as a Connector enum
(Flush / Last / Inner) derived when a node is pushed, so
Node::parent is called exactly once per walk — for the start node, the
only node that can lack a parent. (bca find starts at a matched node
that does have one; it renders as a last child, as before.)

The mirrored dump_metrics / dump_ops walks carried the same per-entry
owned prefix and got the same treatment. Their measured cost is unchanged
on the fixtures tried — a nested-closure chain keeps one stack entry alive
at a time, so the quadratic term needs a tree that is wide and deep —
but the O(depth) copy per rendered line is gone and the three walks no
longer differ in shape.

The last commit (87da2b98) is a follow-up cleanup landed after the issue
write-up: push_children now consumes the child iterator directly and
reverses the new stack tail in place, dropping the reused Vec<Node>
staging buffer so each node is copied once rather than twice.

Measurements

Release binary, the issue's fixture int main(){return ((((…1…))));}:

depth	before	after	peak RSS before	after
1000	84 ms	10 ms	15.5 MB	11.6 MB
2000	284 ms	10 ms	25.4 MB	12.0 MB
4000	1,181 ms	50 ms	66.2 MB	13.0 MB

Wide-and-deep JavaScript fixture (1500 nested functions, two extra
siblings per level), where the parent() term bites hardest:

command	before	after
bca dump	4.00 s	0.05 s
bca metrics	1.33 s	1.29 s

What this does not fix

The emitted byte count, which is what the issue's "~34,000x
amplification" headline measures. A depth-d tree drawing carries d
glyphs of indentation on every line, so output is inherently
O(nodes × depth); 546 MB at depth 8000 is what the format costs. Bounding
the output is a CLI resource-limit decision tracked in #1053 — --depth
is the existing lever. This PR removes the time and the resident memory
spent producing those bytes, not the bytes.

Evidence

• Byte-identity over 300 files spanning src/,
  tests/repositories/pdf.js, tests/repositories/DeepSpeech, and
  tests/repositories/serde, against a binary built from main:
  dump 300/300 and metrics 300/300 byte-identical; ops 300/300 after
  normalising the positional connector and sorting, because bca ops
  output is already nondeterministic run to run on main — filed
  separately as fix(ops): operator/operand order is nondeterministic across runs #1091.
• make pre-commit green end to end (MAKE_EXIT=0) on the tip
  commit: cargo trio, doc warnings, udeps, the markdown/TOML/shell/
  Makefile/Actions lint families, man-page drift, both self-scan tiers,
  and the Python ruff/mypy/pyright/pytest/stubtest stages.
• Every prefix.truncate is covered by a perturbation-verified test.
• No snapshot churn and no submodule bump: the rendered text is unchanged,
  so a clean run leaves no .snap.new.

Not covered by make bench-scaling. Its probes measure the metric
walk, not the dump walk, so no probe exercises this path — the numbers
above are direct release-binary timings on the two fixtures, not harness
output.

Spun out

• fix(ops): operator/operand order is nondeterministic across runs #1091 — bca ops entry order is nondeterministic run to run on
  main, found while establishing byte-parity.

[codecov]

Codecov Report

❌ Patch coverage is 97.75281% with 6 lines in your changes missing coverage. Please review.
✅ Project coverage is 98.04%. Comparing base (edbad0a) to head (87da2b9).
⚠️ Report is 3 commits behind head on main.

Files with missing lines	Patch %	Lines
src/output/dump_metrics.rs	97.16%	0 Missing and 3 partials ⚠️
src/output/dump_ops.rs	95.91%	0 Missing and 2 partials ⚠️
src/output/dump.rs	99.08%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1092      +/-   ##
==========================================
- Coverage   98.06%   98.04%   -0.02%     
==========================================
  Files         270      271       +1     
  Lines       67117    67235     +118     
  Branches    66687    66805     +118     
==========================================
+ Hits        65816    65922     +106     
- Misses        860      871      +11     
- Partials      441      442       +1     

Flag	Coverage Δ
rust	98.03% <97.75%> (-0.02%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
src/output/mod.rs	100.00% <100.00%> (ø)
src/output/dump.rs	92.15% <99.08%> (-3.16%)	⬇️
src/output/dump_ops.rs	82.19% <95.91%> (+2.35%)	⬆️
src/output/dump_metrics.rs	87.10% <97.16%> (+4.31%)	⬆️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.