docpluck 2.4.63__tar.gz → 2.4.65__tar.gz

{docpluck-2.4.63 → docpluck-2.4.65}/CHANGELOG.md

@@ -1,5 +1,63 @@
 # Changelog
 
+## [2.4.65] — 2026-05-22
+
+**Cycle 13 (run 9) — three further normalize_text idempotence fixes.** Post-cycle-12: 11 papers non-idempotent. Cycle 13 packages three independent fixes that together clear 4.
+
+### 1. P1r late re-strip — front-matter metadata leak (4 papers: li-feldman-fox, amp-1, annals-2, xiao-poc-epley)
+
+`_strip_frontmatter_metadata_leaks` matches acknowledgment lines by anchored prefix + keyword guard (`reviewers|editor|feedback|comments|suggestions|insights|helpful`) within 300 chars. pdftotext often line-wraps the acknowledgment BEFORE the guard keyword fires: `We thank the target article's authors - Prof. Craig Fox and Prof. Rebecca Ratner, for being very` (96 chars; no keyword yet). S7/S8 join the continuation; the joined line now contains `helpful in providing us with materials...` — but P1 has already run by then. Pass 2's P1 catches the joined form — non-idempotence + a real missed production strip.
+
+Fix: P1r block at end of `normalize_text`, after H0r and before P0r. Same shape as cycle 7's H0r and cycle 9's P0r — fixed-point re-application of an idempotent line-strip on stabilized line positions.
+
+### 2. Cross-paragraph `=`/`<`/`>` → digit join — same shape as cycle 12 (1 paper: li-feldman-fox additional defect)
+
+A1's `re.sub(r"([=<>])\s*\n\s*([-\d.])", r"\1 \2", t)` uses `\s*` (crosses paragraphs) but runs BEFORE S9 strips header/footer junk. `p =\n\n\x0cFox et al. (2005)...\n\n38\n\n.25, OR = .96, 95%CI [.90, 1.03]` fails on pass 1 (the header text isn't `\s`); S9 strips, leaves `p =\n\n.25`; A1 is over.
+
+Fix: add `re.sub(r"([=<>])\s*\n\s*\n\s*(?=[-.]?\d)", r"\1 ", t)` to the LateJoin block. The lookahead `(?=[-.]?\d)` is the load-bearing constraint — real paragraphs rarely START with a leading dot or `-digit`. Same shape as cycle 12's cross-paragraph `,/;` → `CI/p` joins.
+
+### 3. LABELED CI bracket — intervening-stat-label gate (refines cycle 12; 1 paper: majumder)
+
+Cycle 12's LABELED-bracket discriminator was too permissive. `M = 5.37, SD = 2.01), t(1827) = 1.83, p tukey = .067, d = 0.09, 95% CI [-0.006, 0.18]` has a LABELED `95% CI [...]` that incorrectly paired with `SD = 2.01` (across `t(`, `p tukey =`, `d =` — three INDEPENDENT-statistic labels). The CI is for `d = 0.09`, not for the SD.
+
+Fix: even a LABELED bracket cannot reach back across an independent-stat label. `_INDEPENDENT_STAT_BETWEEN_RE` rejects pairings whose intervening text contains a NEW estimate label (`t`, `F`, `d`, `g`, `OR`, `RR`, `β`, `R²`, `Z`, …) — only variance-family labels (`SD`, `SE`, `M`, `CI`, `%`) are allowed between the candidate token and the labeled CI. efendic's `Mposterior = 20.54, SD=0.04, CI = [-0.61, -0.47]` (only `SD` between) still pairs correctly.
+
+**Impact:** corpus-wide non-idempotency 11 → 7. Broad pytest 1356 pass + 1 known pre-existing B6 fail. Harness Tier-D academic: 0 regressions, 0 new fails (1 still failing — plos-med-1 / B1).
+
+NORMALIZATION_VERSION 1.9.19. Cycle 11/12 contract tests still pass under the refined LABELED-bracket gate.
+
+## [2.4.64] — 2026-05-22
+
+**Cycle 12 (run 9) — three independent normalize_text idempotence fixes.** A 180-doc scan post-cycle-11 found 17 papers still non-idempotent. This cycle packages three independent fixes that together clear 6 of them:
+
+### 1. Final blank-line collapse (5 papers — chan-etal, horsham, lee-feldman, li-feldman-mental-acct, kassambara)
+
+Raw pdftotext output contains form-feed `\x0c` characters at page boundaries. S9's `re.sub(r"\n{3,}", "\n\n", t)` collapses consecutive blank lines, but the form-feed survives upstream stripping into the references region, where R3 (continuation join) processes line-by-line — `"\x0c".strip() == ""` so the form-feed line becomes an empty entry, surrounded by other empty entries. R3 outputs `"\n".join(["...", "", "", "...", ""])` = `\n\n\n\n` (4 newlines). S9's collapse already ran upstream; nothing else collapses. Pass 2 sees the `\n{4}` run and S9 collapses it — non-idempotence.
+
+Fix: add a final `re.sub(r"\n{3,}", "\n\n", t)` right before the H0r/P0r blocks. Any late strip step that empties a line is now safely followed by the collapse, regardless of which step produced the gap.
+
+### 2. Cross-paragraph stat-continuation join (2 papers — korbmacher×2)
+
+A1 (the early stat-line-repair step using `\s*`) crosses paragraph breaks but runs BEFORE S9 strips header/footer noise. A row like
+
+  `r(1798) = -0.27,\n\n472\n\nJournal of Decision Making, Vol. 17...\n\n95% CI [-0.31, ...]`
+
+has so much intervening junk that A1's lookahead fails on pass 1. S9 then strips `472` (page num) and the journal-masthead/page-header (repeated ≥5 times), leaving `-0.27,\n\n95% CI`. A1 is over; LateJoin's A1r uses strict `[ \t]*\n[ \t]*` (single-newline only) and so doesn't fire. Pass 2's A1 sees the now-clean `,\n\n95% CI` and joins — non-idempotence.
+
+Fix: add two paragraph-crossing variants to the LateJoin A1r block, restricted to high-confidence prefixes — `\d+% CI` and `p [<=>]`. No real paragraph STARTS with `95% CI` or `p < .001`, so joining across `\n\n` is safe. The `test_column_bleed_too_many_fragments_ignored` contract is unaffected — its input has no leading `,`/`;`.
+
+### 3. LABELED vs BARE CI bracket discriminator (refines cycle 11)
+
+Cycle 11's proximity gate broke 2 pre-existing tests:
+  - `test_ci_pairing_recovers_body_line`: `Mposterior = 20.54, SD=0.04, CI = [-0.61, -0.47]` — `, SD=` falsely tripped the "new stat label" sentence-break check, blocking the legitimate recovery of `20.54` → `-0.54`.
+  - `test_efendic_table_point_estimates_recovered_via_ci`: efendic's body-line CI recoveries no longer fired.
+
+Fix: discriminate LABELED brackets (`CI = [...]` / `95% CI [...]` / `CI: [...]`) from BARE brackets (`[lo, hi]` alone). LABELED brackets can pair with any candidate token in the row (the chain `M = X, SD = Y, CI = [...]` is all describing the same estimate). BARE brackets retain the strict 30-char + period/semicolon-break proximity gate (catches the majumder false-positive — bare bracket ~50 chars after `2.01`, attached to a different stat). The `_CI_LABEL_PREFIX_RE` looks back ≤8 chars from the `[` for `CI` / `\d+% CI` (with optional `=`/`:`).
+
+**Impact:** corpus-wide non-idempotency 17 → 11 (cycle 12 cleared 6: 5 bibliography-shift + 2 korbmacher; 3 new bibliography cases of the same shape are now caught by the final collapse). Broad pytest 1356 pass + 1 known pre-existing B6 fail. Harness Tier-D academic: 0 regressions, 0 new fails (1 still failing — plos-med-1 / B1).
+
+NORMALIZATION_VERSION 1.9.18. New tests: `test_normalize_collapses_late_blank_line_runs` + `test_late_join_crosses_paragraph_for_stat_continuation`. Cycle 11's tests (`*_proximity_gate_*`) still pass under the LABELED/BARE refinement.
+
 ## [2.4.63] — 2026-05-21
 
 **Cycle 11 (run 9) — `recover_minus_via_ci_pairing` proximity gate.** A 180-doc scan post-cycle-10 found 19 papers still non-idempotent. Among them, 8 (majumder, korbmacher×2, van-boven, chan-feldman-baron, ziano, xiao-poc, amp-1, annals-2) shared a structural defect that ALSO ships in single-pass production: the `_recover_minus_in_record` helper paired every candidate `2X.XX` token with EVERY CI bracket in the same record. A record like `M = 5.37, SD = 2.01), t(1827) = 1.83, p tukey = .067, d = 0.09 [-1.86, 0.04]` contains:

{docpluck-2.4.63 → docpluck-2.4.65}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: docpluck
-Version: 2.4.63
+Version: 2.4.65
 Summary: PDF, DOCX, and HTML text extraction and normalization for academic papers
 Project-URL: Homepage, https://github.com/giladfeldman/docpluck
 Project-URL: Documentation, https://github.com/giladfeldman/docpluck/tree/main/docs

{docpluck-2.4.63 → docpluck-2.4.65}/docpluck/__init__.py

@@ -71,7 +71,7 @@ from .figures import Figure
 from .extract_structured import TABLE_EXTRACTION_VERSION, StructuredResult, extract_pdf_structured
 from .render import render_pdf_to_markdown
 
-__version__ = "2.4.63"
+__version__ = "2.4.65"
 __author__ = "Gilad Feldman"
 __license__ = "MIT"
 

{docpluck-2.4.63 → docpluck-2.4.65}/docpluck/normalize.py

@@ -23,7 +23,7 @@ class NormalizationLevel(str, Enum):
     academic = "academic"
 
 
-NORMALIZATION_VERSION = "1.9.17"
+NORMALIZATION_VERSION = "1.9.19"
 
 
 # ── Mathematical Alphanumeric Symbols de-styling (shared, v2.4.34) ──────────
@@ -1474,35 +1474,64 @@ _CORRUPT_NEG_TOKEN_RE = re.compile(r"(?<![\d.\-])2(\d?\.\d+)\b")
 _TABLE_ROW_RE = re.compile(r"<tr\b.*?</tr>", re.DOTALL | re.IGNORECASE)
 
 
-# Cycle 11 (v2.4.63) — proximity gate for the CI-pairing recovery.
+# Cycle 11 (v2.4.63) / 12 (v2.4.64) — proximity gate for the CI-pairing recovery.
 #
-# In stat reporting the point estimate is IMMEDIATELY followed by its CI:
-# `B = -2.68 [-4.65, -0.68]`. The previous "pair with any bracket in the
-# record" rule false-positives when a record contains an unrelated SD value
-# and a separately-reported CI:
-#   `M = 5.37, SD = 2.01, t(1827) = 1.83, d = 0.09 [-1.86, 0.04]`
-#                  ^^^^                            ^^^^^^^^^^^^^^
-# The `2.01` is a valid SD; the `[-1.86, 0.04]` is the CI for `d = 0.09`.
-# `-0.01` happens to fall inside `[-1.86, 0.04]`, so the old logic
-# recovered `2.01` → `-.01`, corrupting the SD. 8 papers in the corpus
-# (majumder, korbmacher, van-boven, ...) had this defect.
+# In stat reporting a BARE bracket `[lo, hi]` attaches to the IMMEDIATELY-
+# preceding point estimate; a LABELED bracket `CI = [lo, hi]` or
+# `95% CI [lo, hi]` can attach to ANY earlier point estimate on the same
+# row (the SD/SE/df-pair tokens in between are descriptive of the same
+# estimate). The cycle 11 proximity gate treated both as needing strict
+# adjacency, which broke efendic's body-line recovery
+#   `Mposterior = 20.54, SD=0.04, CI = [-0.61, -0.47]`
+# where `, SD=` falsely tripped the "new stat label" sentence-break check.
 #
-# Fix: require the CI bracket to follow the candidate token closely. The
-# bracket's start must be within _CI_PAIR_MAX_GAP chars of the token's
-# end, AND the intervening text must not contain a sentence break (period
-# followed by space, or semicolon, or a comma followed by a new statistic
-# label like ` SD =`/` t(`/` p =`/` d =`).
+# Cycle 12 fix: discriminate LABELED vs BARE brackets.
+#   - LABELED bracket (`CI =`/`95% CI`/`CI:` immediately precedes `[`):
+#     pairs with any candidate token in its record (the old wide rule).
+#   - BARE bracket: pairs ONLY with candidates within 30 chars + no
+#     sentence break (period/semicolon + space — NOT comma + new label,
+#     because stat-row labels are comma-separated by convention).
+#
+# This keeps the majumder fix (bare bracket far from `2.01`) AND
+# preserves efendic-style labeled CIs that pair across SD/SE annotations.
 _CI_PAIR_MAX_GAP = 30
-_SENTENCE_BREAK_RE = re.compile(
-    r"[.;]\s|,\s+(?:SD|SE|t|F|p|d|g|η|χ|r|R²|β|B|γ|R|N|M|Q|Z)\s*[=(]",
-    re.IGNORECASE,
+# Bare-bracket sentence break: only period/semicolon + space.  A comma is
+# NOT a break because stat rows are comma-separated.  The majumder false-
+# positive is now caught by the per-bracket proximity check (the bare
+# bracket sits ~50 chars after `2.01` — beyond _CI_PAIR_MAX_GAP).
+_SENTENCE_BREAK_RE = re.compile(r"[.;]\s")
+# A bracket is "labeled" when prefixed by `CI`, `95 % CI`, or similar
+# directly before the opening `[`. Allow optional whitespace and an `=` /
+# `:` between the label and the bracket.
+_CI_LABEL_PREFIX_RE = re.compile(r"(?:\bCI|\b\d+\s*%\s*CI)\s*[=:]?\s*$", re.IGNORECASE)
+# Cycle 13 (v2.4.65) — even a LABELED CI cannot reach back ACROSS an
+# independent-test-statistic label. The discriminator: between the
+# candidate token and the labeled bracket, allow ONLY variance-family
+# labels (SD, SE, M, Mdn, Var, CI, 95% CI itself, %), reject anything
+# that introduces a NEW estimate (t, F, p, d, g, η, χ, r, R², β, OR, RR,
+# HR, B, Z, Q).
+#
+# Why: `Mposterior = 20.54, SD=0.04, CI = [-0.61, -0.47]` (efendic) has
+# only SD between the candidate and the CI — same estimate, paired OK.
+# `M = 5.37, SD = 2.01, t(1827) = 1.83, p tukey = .067, d = 0.09, 95% CI
+# [-0.006, 0.18]` (majumder) has t, p, d — three independent estimates —
+# between `2.01` and the CI; the CI is for `d`, not `2.01`. Reject.
+_INDEPENDENT_STAT_BETWEEN_RE = re.compile(
+    r"(?:^|[,;\s])\s*"
+    r"(?:t|F|d|g|R|R²|β|γ|B|OR|RR|HR|H|Q|Z|f|n|η|χ|η²|χ²|r|"
+    r"p\s+tukey|p\s+holm|p\s+bonf(?:erroni)?|p\s+adj|"
+    r"\bp(?:\s*[=<>]))"
+    r"\s*[=(\(]",
 )
 
 
 def _recover_minus_in_record(record: str) -> str:
     """Recover '2X.XX' tokens in a single record (a table row or a text line)
     by pairing each with a CI bracket present in the same record."""
-    brackets: list[tuple[float, float, tuple[int, int]]] = []
+    # Each entry: (lo, hi, (bs, be), is_labeled). `is_labeled` is True when
+    # the bracket is prefixed by `CI`/`95% CI`/etc. — see cycle 12 notes
+    # at _CI_LABEL_PREFIX_RE.
+    brackets: list[tuple[float, float, tuple[int, int], bool]] = []
     for m in _CI_PAIR_BRACKET_RE.finditer(record):
         try:
             lo, hi = float(m.group(1)), float(m.group(2))
@@ -1510,13 +1539,17 @@ def _recover_minus_in_record(record: str) -> str:
             continue
         if lo > hi:
             continue  # not a well-formed interval
-        brackets.append((lo, hi, m.span()))
+        # Look back ≤8 chars for a `CI` / `95 % CI` label.
+        bs, be = m.span()
+        prefix = record[max(0, bs - 8): bs]
+        is_labeled = bool(_CI_LABEL_PREFIX_RE.search(prefix))
+        brackets.append((lo, hi, (bs, be), is_labeled))
     if not brackets:
         return record
 
     def _sub(m: "re.Match[str]") -> str:
         # Never touch a token that lies inside a bracket span (a CI bound).
-        for _lo, _hi, (bs, be) in brackets:
+        for _lo, _hi, (bs, be), _lab in brackets:
             if bs <= m.start() < be:
                 return m.group(0)
         frac = m.group(1)
@@ -1525,22 +1558,33 @@ def _recover_minus_in_record(record: str) -> str:
             recovered = float("-" + frac)
         except ValueError:
             return m.group(0)
-        # Cycle 11 proximity gate: only consider brackets that closely
-        # follow the token, with no sentence break or new stat label in
-        # between. Pick the NEAREST eligible bracket (the one that would
-        # canonically pair with the point estimate).
+        # Cycle 12: pick the NEAREST bracket whose pairing rules accept this
+        # token. LABELED brackets accept any candidate in the record (legacy
+        # wide rule — efendic body line `Mposterior = 20.54, SD=0.04,
+        # CI = [-0.61, -0.47]` is the canonical case). BARE brackets only
+        # accept the immediately-preceding stat (within 30 chars, no
+        # sentence break) — this is what blocks the majumder false-positive
+        # `M = 5.37, SD = 2.01, t = ..., d = 0.09 [-1.86, 0.04]`.
         token_end = m.end()
         nearest = None
         nearest_dist = None
-        for lo, hi, (bs, be) in brackets:
+        for lo, hi, (bs, be), is_labeled in brackets:
             if bs < token_end:
-                continue  # bracket precedes the token — not its CI
-            gap = bs - token_end
-            if gap > _CI_PAIR_MAX_GAP:
                 continue
+            gap = bs - token_end
             intervening = record[token_end:bs]
-            if _SENTENCE_BREAK_RE.search(intervening):
-                continue
+            if is_labeled:
+                # Labeled bracket: relaxed proximity, but still reject if
+                # an independent-stat label intervenes. The label gates the
+                # pairing to the variance-family (SD/SE/M/CI/%) of the
+                # SAME estimate. See _INDEPENDENT_STAT_BETWEEN_RE notes.
+                if _INDEPENDENT_STAT_BETWEEN_RE.search(intervening):
+                    continue
+            else:
+                if gap > _CI_PAIR_MAX_GAP:
+                    continue
+                if _SENTENCE_BREAK_RE.search(intervening):
+                    continue
             if nearest_dist is None or gap < nearest_dist:
                 nearest = (lo, hi)
                 nearest_dist = gap
@@ -2649,6 +2693,31 @@ def normalize_text(
         t = re.sub(r"([=<>])[ \t]*\n[ \t]*(?=[-\d.])", r"\1 ", t)
         t = re.sub(r"([,;])[ \t]*\n[ \t]*(?=p\s*[<=>])", r"\1 ", t)
         t = re.sub(r"([,;])[ \t]*\n[ \t]*(?=\d+%\s*CI)", r"\1 ", t)
+        # Cycle 12 (v2.4.64) — cross-paragraph stat-continuation join.
+        # A1 (which uses `\s*` and so crosses paragraph breaks) runs BEFORE
+        # S9 strips header/footer lines. So a stat row like
+        #   `r(1798) = -0.27,\n\n472\n\nJournal of Decision Making, ...\n\n95% CI [-0.31, ...]`
+        # has so much intervening junk that A1's lookahead fails on pass 1;
+        # only after S9 strips the junk (producing `,\n\n95% CI`) can the
+        # join happen, and that's pass 2. The two patterns below are the
+        # paragraph-crossing variants of the comma-to-stat-continuation
+        # patterns above — restricted to the high-confidence prefixes
+        # `\d+% CI` and `p [<=>]` because no real paragraph STARTS with
+        # those tokens (test_column_bleed_too_many_fragments_ignored is
+        # unaffected — its input has no leading `,`/`;`).
+        # Clears korbmacher (2 papers) from the non-idempotent set.
+        t = re.sub(r"([,;])\s*\n\s*\n\s*(?=\d+%\s*CI)", r"\1 ", t)
+        t = re.sub(r"([,;])\s*\n\s*\n\s*(?=p\s*[<=>])", r"\1 ", t)
+        # Cycle 13 (v2.4.65) — same shape, applied to `=/<>` → digit/dot
+        # continuations. li-feldman-fox has `p =\n\n\x0cFox et al. (2005)...
+        # \n\n38\n\n.25, OR = .96, 95%CI [.90, 1.03])` where A1's
+        # `([=<>])\s*\n\s*([-\d.])` pattern fails on pass 1 (the journal-
+        # header text isn't `\s`); S9 strips the header + page number,
+        # leaving `p =\n\n.25` — but A1 is over. Pass 2 joins on the
+        # cleaned form. The lookahead `(?=[-\d.])` is the load-bearing
+        # constraint — real paragraphs rarely START with a leading dot
+        # or `-digit`.
+        t = re.sub(r"([=<>])\s*\n\s*\n\s*(?=[-.]?\d)", r"\1 ", t)
     report._track("LateJoin_line_break_rejoin", before, t, "late_line_joins")
 
     # ── H0r: header-banner re-strip on stabilized line positions ─────────
@@ -2669,6 +2738,54 @@ def normalize_text(
         t = _restripped
     report._track("H0r_header_banner_restrip", before, t, "header_banners_restripped")
 
+    # ── Final blank-line collapse ────────────────────────────────────────
+    # S9 enforces `re.sub(r"\n{3,}", "\n\n", t)` once near the top of the
+    # pipeline. Later steps that REMOVE non-blank content can leave blank
+    # gaps that S9's earlier collapse no longer reaches:
+    #
+    #   - R3 (refs-section continuation join) walks the refs span line by
+    #     line. A bare form-feed `\x0c` (pdftotext page-break) between two
+    #     blank lines becomes `"".strip() == ""` and is preserved as a blank
+    #     entry; R3 outputs three consecutive blank entries surrounded by
+    #     `"\n".join(...)` — `\n\n\n\n`. Pass 1 leaves this; pass 2's S9
+    #     collapses it, producing the bibliography-shift non-idempotence
+    #     (cycle 12 — 5 papers: chan-etal, horsham, lee-feldman,
+    #     li-feldman-mental, + 1 incidental).
+    #   - Same pattern for any late strip step that empties a line without
+    #     re-collapsing.
+    #
+    # Add the collapse here so the function is idempotent regardless of
+    # which late step produced the blank-line run.
+    t = re.sub(r"\n{3,}", "\n\n", t)
+
+    # ── P1r: front-matter metadata-leak re-strip on stabilized lines ─────
+    # Same shape as H0r and P0r. P1's `_strip_frontmatter_metadata_leaks`
+    # matches an acknowledgment-style line by ANCHORED prefix + a keyword
+    # check within the first 300 chars (e.g. `^We\s+thank...reviewers|
+    # editor|feedback|comments|suggestions|insights|helpful`). pdftotext
+    # often line-wraps the acknowledgment before the keyword fires (e.g.
+    # `We thank the target article's authors - Prof. Craig Fox and Prof.
+    # Rebecca Ratner, for being very` — the raw line stops before
+    # `helpful`). S7/S8 join the continuation; the joined line now contains
+    # the keyword, but P1 has already run by then. Pass 2's P1 catches the
+    # joined form and strips — non-idempotence + a missed production strip.
+    #
+    # Re-running here on the post-LateJoin line positions catches every
+    # form (the original short line where the keyword was already in
+    # window, AND the post-join long line where it's only in window after
+    # the join).
+    #
+    # Cycle 13 (v2.4.65) — clears li-feldman-fox + amp-1 + annals-2 +
+    # xiao-poc-epley (4 acknowledgment-block papers) from the
+    # non-idempotent set.
+    before = t
+    while True:
+        _restripped = _strip_frontmatter_metadata_leaks(t)
+        if _restripped == t:
+            break
+        t = _restripped
+    report._track("P1r_frontmatter_leak_restrip", before, t, "frontmatter_leaks_restripped")
+
     # ── P0r: page-footer-line re-strip on stabilized line positions ──────
     # Same shape as H0r, applied to P0's anchored ^...$ patterns. P0 runs
     # near the top of the pipeline, where some P0-targeted lines are still

{docpluck-2.4.63 → docpluck-2.4.65}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "docpluck"
-version = "2.4.63"
+version = "2.4.65"
 description = "PDF, DOCX, and HTML text extraction and normalization for academic papers"
 readme = "docs/README.md"
 requires-python = ">=3.10"

{docpluck-2.4.63 → docpluck-2.4.65}/tests/test_normalize_idempotent_real_pdf.py

@@ -234,6 +234,45 @@ def test_s9_4digit_pattern_a_still_strips_isolated_page_numbers():
     )
 
 
+def test_normalize_collapses_late_blank_line_runs():
+    """Cycle 12: a late strip step that empties a line (e.g. R3 stripping a
+    form-feed `\\x0c` between two blank lines) leaves a `\\n{3+}` run that
+    S9's earlier collapse no longer reaches. The final collapse at the end
+    of normalize_text catches it."""
+    # Simulate: paragraph + form-feed + paragraph (pdftotext page-break case)
+    text = "First paragraph ending here.\n\n\x0c\n\nSecond paragraph begins."
+    out, _ = normalize_text(text, NormalizationLevel.academic)
+    # Should produce one paragraph break, not two
+    assert "\n\n\n" not in out, f"normalize_text left a \\n{{3+}} run: {out!r}"
+    assert "First paragraph" in out
+    assert "Second paragraph" in out
+
+
+def test_late_join_crosses_paragraph_for_stat_continuation():
+    """Cycle 12: a comma/semicolon followed by a paragraph break and a
+    high-confidence stat-continuation token (95% CI / p [<=>]) is a
+    serializer artifact — joined on pass 1. Pre-cycle-12, only pass 2
+    joined it (after S9 stripped the intervening header/footer noise).
+
+    Defends against the corpus-wide korbmacher pattern where a regression-
+    coefficient row was broken by a per-page header insertion."""
+    # The full korbmacher pre-S9 pattern is a couple of headers between;
+    # post-S9 the input to LateJoin is just `,\n\n95% CI`.
+    text = "r(1798) = -0.27,\n\n95% CI [-0.31, -0.22]"
+    out, _ = normalize_text(text, NormalizationLevel.academic)
+    assert "-0.27, 95% CI" in out, f"cross-paragraph stat join failed: {out!r}"
+
+    # Same for p-value continuation
+    text2 = "t(23) = 2.34,\n\np < .001, d = 0.45"
+    out2, _ = normalize_text(text2, NormalizationLevel.academic)
+    assert "2.34, p < .001" in out2 or "p < .001" in out2.replace("\n\n", " ")
+
+    # The column-bleed contract is NOT broken — its input has no leading `,;`.
+    cb = "p\n01\n02\n03\n04\n05\n= .05"
+    out_cb, _ = normalize_text(cb, NormalizationLevel.academic)
+    assert "p = .05" not in out_cb, "column-bleed test contract broken by cycle 12"
+
+
 def test_recover_minus_proximity_gate_rejects_distant_unrelated_brackets():
     """Cycle 11: a stat-table row that mixes an unrelated SD value with a
     separately-reported CI bracket must NOT have the SD recovered as a