Fix two quadratic walks: tail_reading's fixed point and the particle chain's title test

[derek73]

Closes #558, closes #559. Both are found by the _pipeline/ sweep for #553, and both are performance fixes that move no field, report or view.

What changed

#559, _group.py chain(). The ambiguous-particle emitter asked all(is_leading_title(...) for x in range(k)) at every chain site, which costs leading titles × sites. merge(k, j) changes only indices from k on, and k only grows, so every piece behind k is final. A forward-only titled cursor therefore answers the same question with one walk per chain.

#558, _pieces.py tail_reading. The S2-peel / H5-chain fixed point re-ran peel_trailing over the whole walk once per title the chain took. A pass now resumes the walk at the splice, over rest as written:

• The pieces in front of the splice are the ones a fresh walk meets.
• credential_anchors is a front-to-back pass, so its carried values still hold for those positions.

Two of the walk's tests read position counts of pieces already peeled, and the splice lowers those counts. So a pass resumes only where neither count can move:

• kept >= 2: every carried member still counts ≥ 3 for the acronym fork's words-to-spare test.
• behind_count >= 2: the numeral's last-pair test reads the same pair.

Otherwise the pass walks afresh. A pick that the stopping title made is dropped. The splice is never materialized between passes; its runs are joined once at the end, so no C-level copy goes quadratic either.

The first pass is unchanged (end= replaces a slice), so ordinary names pay no new frame and the _CALL_BASELINE band holds.

Evidence that nothing moves

• Oracle on every tail_reading call. I wrapped every call site to run the old re-peeling loop next to the new one and compared (rest, titled, names, numeral, picks). The inputs were the differential corpora plus 150,000 fuzzed names under four configurations. One configuration is a lexicon putting ma/ba/sa in the titles, which makes a stopping title an ambiguous-acronym pick. Result: 698,009 calls, 108,453 through the resume path, all identical.
• Negative controls. Weakening each condition fails the oracle:
  • kept >= 1 fails on phd rev. ma jr..
  • behind_count >= 1 fails on John Smith V Prof. VI.
  • Never dropping the stopper pick fails only under the overlap lexicon, because in the default vocabulary no title is an ambiguous acronym.
• Whole-parse diff against master. repr, ambiguities and initials() were compared over 121,745 names × 3 configurations: byte-identical, with nameparser.__file__ asserted on each side. An independent review fuzz (72,663 names, plus a 36,012-name chain fuzz with 13,279 particle-or-given reports) also found no difference.
• Differential gate. Exit 0 at all five baselines (1.4.0, 2.0.0, 2.1.0, 2.2.0, 2.3.0). The output is identical to master, so the classified summaries are too.
• Checks. Suite 10616 passed, plus mypy and ruff.

Guards

_REREAD_SHAPES in tests/v2/test_benchmark.py counts the frames of the one function each defect re-ran, at 8 units against 32:

shape	fixed	at fc682e3
tail (listed_lean)	3.67x	14.67x
clause (listed_lean)	3.67x	14.67x
chain (is_leading_title)	3.77x	11.00x

Each row fails against a copy reverting only its own half of the fix, and passes against one reverting only the other.

These are frame-counted rather than _PREFIXED_SHAPES clock rows, which is what #558 suggested. The cost is Python-level, so a frame count is deterministic and keeps running under a line tracer, while that table exists for C-level costs a frame can't see (#553). AGENTS.md's perf gotcha now names the new table.

test_a_title_the_fixed_point_splices_out_keeps_no_peel_pick (tests/v2/pipeline/test_pieces.py) covers the stopper-pick drop, the one diff line no other test reached. The drop is reachable only through a caller's lexicon, so the test adds ma to the titles. Its recorded control: without the drop, John Smith Ma. PhD Jr. keeps its fields but reports suffix-or-name on the title Ma..

Release log

Two bullets, measured on Python 3.11 through HumanName, 400 → 1,600 units:

shape	2.0.0	2.2.0	2.3.0	this branch
#558 tail	4.2x	4.1x	12.9x	4.1x
#559 chain	12.5x	12.2x	12.2x	4.1x

The #558 maiden-clause shape was never released, so no bullet claims it.

Left alone

decisions.md's 2026-09-09 fixed-point entry says John Prof. MA Prof. "reads … family MA". Since #289, mixed-case MA takes the credential, so that reading is now one-case only. The entry is a dated record, so I've left it as it was; the live docstring next to it now names the one-case pair.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.87%. Comparing base (fc682e3) to head (7a4fa38).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #572   +/-   ##
=======================================
  Coverage   98.87%   98.87%           
=======================================
  Files          45       45           
  Lines        3990     4017   +27     
=======================================
+ Hits         3945     3972   +27     
  Misses         45       45           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.